Digital Edition

Digital Editions of selected Intel Press books are in

addition to and complement the printed books.

Click the icon to access information on other essential

books for Developers and IT Professionals
Visit our website at www.intel.com/intelpress
Multi-Core Programming
Increasing Performance through Software
Multi-threading

Shameem Akhter
Jason Roberts

Intel
PRESS
Copyright 2006 Intel Corporation. All rights reserved.
ISBN 0-9764832-4-6

This publication is designed to provide accurate and authoritative information in regard to the
subject matter covered. It is sold with the understanding that the publisher is not engaged in
professional services. If professional advice or other expert assistance is required, the services
of a competent professional person should be sought.

Intel Corporation may have patents or pending patent applications, trademarks, copyrights, or
other intellectual property rights that relate to the presented subject matter. The furnishing of
documents and other materials and information does not provide any license, express or
implied, by estoppel or otherwise, to any such patents, trademarks, copyrights, or other
intellectual property rights.

Intel may make changes to specifications, product descriptions, and plans at any time, without
notice.

Fictitious names of companies, products, people, characters, and/or data mentioned herein are
not intended to represent any real individual, company, product, or event.

Intel products are not intended for use in medical, life saving, life sustaining, critical control or
safety systems, or in nuclear facility applications.

Intel, the Intel logo, Celeron, Intel Centrino, Intel NetBurst, Intel Xeon, Itanium, Pentium, MMX,
and VTune are trademarks or registered trademarks of Intel Corporation or its subsidiaries in
the United States and other countries.

Other names and brands may be claimed as the property of others.

This book is printed on acid-free paper.

Publisher: Richard Bowles

Editor: David J. Clark
Managing Editor: David B. Spencer
Content Architect: Stuart Goldstein
Text Design & Composition: Interactive Composition Corporation
Graphic Art: Kirsten Foote (illustrations), Ted Cyrek (cover)
Library of Congress Cataloging in Publication Data:
Printed in the United States of America
10 9 8 7 6 5 4 3 2 1
First printing, April 2006
To my beloved wife Munny and my parents.
S.A.

To my mother.
J.R.
 Contents
Preface xi
Chapter 1 Introduction to Multi-Core Architecture 1
Motivation for Concurrency in Software 2
Parallel Computing Platforms 5
Parallel Computing in Microprocessors 7
Differentiating Multi-Core Architectures from Hyper-Threading
Technology 10
Multi-threading on Single-Core versus Multi-Core Platforms 11
Understanding Performance 13
Amdahls Law 14
Growing Returns: Gustafsons Law 18
Key Points 19

Chapter 2 System Overview of Threading 21

Defining Threads 22
System View of Threads 22
Threading above the Operating System 23
Threads inside the OS 26
Threads inside the Hardware 29
What Happens When a Thread Is Created 30
Application Programming Models and Threading 32
Virtual Environment: VMs and Platforms 33
Runtime Virtualization 33

v
vi Multi-Core Programming

System Virtualization 33
Key Points 35

Chapter 3 Fundamental Concepts of Parallel

Programming 37
Designing for Threads 37
Task Decomposition 38
Data Decomposition 39
Data Flow Decomposition 40
Implications of Different Decompositions 41
Challenges Youll Face 42
Parallel Programming Patterns 42
A Motivating Problem: Error Diffusion 45
Analysis of the Error Diffusion Algorithm 48
An Alternate Approach: Parallel Error Diffusion 48
Other Alternatives 50
Key Points 51

Chapter 4 Threading and Parallel Programming

Constructs 53
Synchronization 53
Critical Sections 56
Deadlock 57
Synchronization Primitives 59
Semaphores 60
Locks 63
Condition Variables 66
Messages 68
Flow Control-based Concepts 71
Fence 71
Barrier 72
Implementation-dependent Threading Features 73
Key Points 74

Chapter 5 Threading APIs 75

Threading APIs for Microsoft Windows 75
Win32/MFC Thread APIs 75
Threading APIs for Microsoft .NET Framework 107
Creating Threads 107
 Contents vii

Managing Threads 110

Thread Pools 112
Thread Synchronization 117
POSIX Threads 120
Creating Threads 120
Managing Threads 122
Thread Synchronization 123
Signaling 124
Compilation and Linking 132
Key Points 132

Chapter 6 OpenMP: A Portable Solution for

Threading 135
Challenges in Threading a Loop 137
Loop-carried Dependence 137
Data-race Conditions 140
Managing Shared and Private Data 141
Loop Scheduling and Partioning 143
Effective Use of Reductions 147
Minimizing Threading Overhead 149
Work-sharing Sections 151
Performance-oriented Programming 152
Using Barrier and Nowait 152
Interleaving Single-thread and Multi-thread Execution 154
Data Copy-in and Copy-out 155
Protecting Updates of Shared Variables 157
Intel Taskqueuing Extension to OpenMP 160
OpenMP Library Functions 162
OpenMP Environment Variables 163
Compilation 164
Debugging 165
Performance 167
Key Points 169

Chapter 7 Solutions to Common Parallel Programming

Problems 171
Too Many Threads 171
Data Races, Deadlocks, and Live Locks 174
viii Multi-Core Programming

Deadlock 177
Heavily Contended Locks 181
Priority Inversion 181
Solutions for Heavily Contended Locks 183
Non-blocking Algorithms 186
ABA Problem 188
Cache Line Ping-ponging 190
Memory Reclamation Problem 190
Recommendations 191
Thread-safe Functions and Libraries 192
Memory Issues 193
Bandwidth 193
Working in the Cache 194
Memory Contention 197
Cache-related Issues 200
False Sharing 200
Memory Consistency 204
Current IA-32 Architecture 204
Itanium Architecture 207
High-level Languages 210
Avoiding Pipeline Stalls on IA-32 211
Data Organization for High Performance 212
Key Points 213

Chapter 8 Multi-threaded Debugging Techniques 215

General Debug Techniques 215
Designing with Debugging in Mind 216
Extending your ApplicationUsing Trace Buffers 219
Debugging Multi-threaded Applications in Windows 224
Threads Window 225
Tracepoints 225
Breakpoint Filters 226
Naming Threads 227
Putting it All Together 228
Multi-threaded Debugging Using GDB 232
Notification on Thread Creation 233
Getting a List of All Threads in the Application 233
Setting Thread-specific Breakpoints 233
 Contents ix

Switching between Threads 235

Applying a Command to a Group of Threads 235
Key Points 236

Chapter 9 Single-Core Processor Fundamentals 237

Processor Architecture Fundamentals 237
Comparing Superscalar and EPIC Architecture 245
Key Points 246

Chapter 10 Threading on Intel Multi-Core Processors 247

Hardware-based Threading 247
Threading from Intel 251
Hyper-Threading Technology 252
Difference between Multiprocessor and Hyper-Threading
Technology 254
Hyper-Threading Technology Architecture 254
Multi-Core Processors 257
Architectural Details 257
Comparison between Multiprocessors and Multi-Core
Processors 260
Multi-Core for Itanium Architecture 261
Multiple Processor Interaction 266
Inter-Processor Communication and Multi-threaded
Programming 266
Power Consumption 268
Power Metrics 268
Reducing Power Consumption 270
Beyond Multi-Core Processor Architecture 271
Key Points 272

Chapter 11 Intel Software Development Products 275

Overview 275
Investigate 276
Create/Express 276
Debugging 277
Tuning 277
Intel Thread Checker 277
How It Works 278
Usage Tips 280
x Multi-Core Programming

Using Intel Thread Checker with OpenMP 281

Intel Compilers 281
OpenMP 282
Software-based Speculative Precomputation 286
Compiler Optimization and Cache Optimization 287
Intel Debugger 288
Intel Libraries 289
Intel Math Kernel Library 289
Intel Integrated Performance Primitives 290
Parallel Program Issues When Using Parallel Libraries 290
The Future 291
Intel Threading Building Blocks 292
Intel VTune Performance Analyzer 292
Find the Hotspot 293
Using Call Graph for Finding a Threading Point 294
Check the Load Balancing 295
Intel Thread Profiler 295
MPI Programming 296
Intel Support for MPI 297
Key Points 300

Glossary 303
References 317
Index 323
 Preface
By now, most technology professionals have heard of the radical
transformation taking place in the way that modern computing platforms are
being designed. Intel, IBM, Sun, and AMD have all introduced microprocessors
that have multiple execution cores on a single chip. In 2005, consumers had
the opportunity to purchase desktop platforms, servers, and game consoles
that were powered by CPUs that had multiple execution cores. Future
product roadmaps show that this is only the beginning; rather than racing
towards being the first to 10 gigahertz, semiconductor manufacturers are now
working towards the goal of leading the industry in the number of execution
cores integrated onto a single die. In the future, computing platforms,
whether they are desktop, mobile, server, or specialized embedded platforms
are most likely to be multi-core in nature.
The fact that the hardware industry is moving in this direction
presents new opportunities for software developers. Previous hardware
platforms presented a sequential programming model to the
programmer. Operating systems and other system software simulated
multitasking environments by exploiting the speed, or lack thereof, of
human perception. As a result, multi-threading was an effective illusion.
With modern multi-core architectures, developers are now presented
with a truly parallel computing platform. This affords software
developers a great deal more power in terms of the ways that they design
and implement their software. In this book, well take a look at a variety
of topics that are relevant to writing software for multi-core platforms.

Intended Audience
Our primary objective is to provide the material software developers need
to implement software effectively and efficiently on parallel hardware

xi
xii Multi-Core Programming

platforms. These platforms include multi-core processors and processors

that use simultaneous multi-threading techniques, such as Hyper-Threading
Technology (HT Technology). This book will focus on programming
techniques that allow the developer to exploit the capabilities provided by
the underlying hardware platform. We hope to reach as broad an audience
as possible. In an effort to accomplish this, weve included introductory
material on basic threading concepts for those new to multi-threaded
programming. However, parallel programming is a complex subject, one
with many different approaches and philosophies. Our intent, and
challenge, is to provide a comprehensive discussion of the hardware
capabilities provided by multi-core processors and platforms using
simultaneous multi-threading techniques without getting lost in the
different academic arguments about whether or not a particular approach
is the optimal solution to the problem of writing reliable, easy to maintain,
parallel programs. References to the definitive works on these topics, as
they are introduced in the text, are included for readers who want to
explore these topics in more detail.
We assume that the reader is an experienced programmer with little to
no background in multi-threaded programming techniques. This may be an
overly cautious assumption; many developers reading this text have
probably used threads when writing applications designed to run on a
single-core platform, are familiar with the basic principles of multi-threaded
programming, and have knowledge of at least one threading API. However,
its important to remember the key differences when writing applications
targeting single-core, multi-core, and platforms with simultaneous multi-
threading technology. These differences are discussed in this book. For this
reason, we have chosen to start from the beginning, highlighting these
differences in the appropriate sections. The authors do assume that the
reader is familiar with at least one high-level language, preferably C/C++.
However, many of the suggestions made in this book apply equally to
languages such as Java and Perl. We hope this approach accomplishes our
goal of making this book relevant to as wide an audience as possible.

About This Book

This book is organized into three major sections. The first section
(Chapters 14) presents an introduction to software threading. This
section includes background material on why chipmakers have shifted to
multi-core architectures, how threads work, how to measure the
performance improvements achieved by a particular threading
 Preface xiii

implementation, programming paradigms for parallel hardware

platforms, and abstract data types used when working with threads. After
completing these chapters, the reader should have a sense of the reasons
why hardware platforms are evolving in the way that they are and
understand the basic principles required to write parallel programs.
The next section of the book (Chapters 5 and 6) discusses common
programming APIs for writing parallel programs. We look at three
programming interfaces: Microsofts APIs for Win32, MFC, and .NET;
POSIX Threads; and OpenMP. We recognize that a large number of
different APIs and programming models are available to developers.
However, given the constraints of time and space, we have chosen a
representative sample of the most commonly used APIs today.
The third and final section is a collection of topics related to multi-
core programming. Chapter 7 discusses common parallel programming
problems and how to solve them. Chapter 8 examines the topic of
debugging multi-threaded implementations. Chapter 9 provides an
introduction or review of hardware fundamentals, and Chapter 10
follows this up with an in-depth discussion of multi-core processors at
the hardware level. In Chapter 11, we talk about the software tools
developed by Intel that help software developers write, debug, and
profile multi-threaded applications.

Finally, it should be noted that all of the Windows based samples
provided with this book were compiled and built with Microsofts Visual

Studio 2005. These applications were tested using Microsoft XP with

Service Pack 2 installed. For Linux , the gcc compiler was used and the
examples were tested on Linux 2.6. All OpenMP examples were

compiled using the latest Intel C++ Compiler. For the code samples,
updates, errata, and additional material, please visit the books Web site:
http://www.intel.com/intelpress/mcp.

Intel Software Development Products
As youll see throughout the text, and especially in Chapter 11, Intel
provides more than just multi-core processors. In addition to the
hardware platform, Intel has a number of resources for software
developers, including a comprehensive tool suite for threading that
includes:
Intel C++ and Fortran compilers, which support multi-threading
by providing OpenMP and automatic parallelization support
xiv Multi-Core Programming

Intel Math Kernel Library and Intel Integrated Performance

Primitives that are threaded via OpenMP
Intel VTune Performance Analyzer, which can be used to
monitor processor events related to multi-threaded performance
Intel Thread Checker and the Intel Debugger, which help debug
common multi-threaded problems like deadlocks
Intel Thread Profiler, which helps developers optimize OpenMP,
Microsoft Windows, and POSIX-based multi-threaded applications
In addition to tools, the Intel Software Network is focused on working
with software vendors around the world to help develop and deploy
production applications. The Intel Software Network consists of a
number of different resources. One of these resources is a detailed
knowledge base of whitepapers and articles written by Intel architects
that share key insights on technology, including optimal threading
techniques. The Intel Software Network also includes user discussion
forums where you can interact with Intel engineers to discuss all things
related to technology. The Intel Software College provides training
courses in a variety of formats, including Webcasts, online training, and
classroom-based training. These classes discuss a wide variety of topics
including multi-threaded programming. Intel Solution Services provides
consulting services for companies looking to get expert advice on a
challenging technical issue.
To start exploring the online resources available to developers
targeting multi-core platforms, visit Intels multi-core homepage at:
http://www.intel.com/multi-core/.

Acknowledgements
This book is the culmination of the efforts of a number of talented
individuals. There are many people that need to be recognized. Wed like
to start off with the list of contributors that developed content for this

book. Chapter 6, OpenMP : A Portable Solution for Threading was
written by Xinmin Tian. Chapter 7, Solutions to Common Parallel
Programming Problems, was written by Arch Robison. Finally, James
Reinders, with contributions by Eric Moore and Gordon Saladino,
developed Chapter 11, Intel Software Development Products. Other
contributors who developed material for this book include: Sergey
Zheltov, Stanislav Bratanov, Eugene Gorbatov, and Cameron McNairy.
 Preface xv

No Intel Press book is published without peer review. Wed like to

thank all the reviewers for identifying errors and for providing valuable
insight and encouragement along the way. Without their help, this book
would not have been a success. From Intel, these individuals
participated, at one time or another, in the review of this project: Andy
Anderson, Jeff Austin, Ernie Brickell, Daniel Brown, Doris Burrill, Stephen
Chao, Brad Corrion, Jim Cownie, David Grawrock, Gerard Hartnett,
William Holt, James Howard, Bob Kuhn, Tim Mattson, My-Hanh Nguyen,
Paul Petersen, James Reinders, Arch Robison, Thom Sawicki, Sanjiv Shah,
Xinmin Tian, Kimberly Turner, Richard Uhlig, Greg Welch, Paul Work,
and Sergey Zheltov. Other reviewers included Andrew Binstock of Pacific
Data Works, LLC, Thomas Kinsman of Eastman Kodak, and Hari Kalva,
Assistant Professor at Florida Atlantic University.
Finally, wed like to thank the team from Intel Press. Stuart Goldstein
was the content architect for this projecthis steady guidance and
ability to find resources as we needed them kept this book on track.
David Clark was the editor on the project, and helped take our
engineering prose and turn it into a real book. David Spencer was
involved in the early stages of this project, getting us started on the right
foot.
For anyone that we may have missed, please accept our apologies.
 Chapter 1
Introduction
to Multi-Core
Architecture
n 1945, mathematician John von Neumann, with the aid of J. Presper
I Eckert and John Mauchly, wrote a memo proposing the creation of an
Electronic Discrete Variable Automatic Computer, more famously known
as the EDVAC. In this paper, von Neumann suggested the stored-program
model of computing. In the von Neumann architecture, a program is a
sequence of instructions stored sequentially in the computers memory.
The programs instructions are executed one after the other in a linear,
single-threaded fashion.
As time went on, advancements in mainframe technology expanded
upon the ideas presented by von Neumann. The 1960s saw the advent of
time-sharing operating systems. Run on large mainframe computers,
these operating systems first introduced the concept of concurrent
program execution. Multiple users could access a single mainframe
computer simultaneously and submit jobs for processing. From the
programs perspective, it was the only process in the system. The operating
system handled the details of allocating CPU time for each individual
program. At this time, concurrency existed at the process level, and the
job of task switching was left to the systems programmer.
In the early days of personal computing, personal computers, or PCs,
were standalone devices with simple, single-user operating systems. Only
one program would run at a time. User interaction occurred via simple
text based interfaces. Programs followed the standard model of straight-
line instruction execution proposed by the von Neumann architecture.
Over time, however, the exponential growth in computing performance

1
2 Multi-Core Programming

quickly led to more sophisticated computing platforms. Operating system

vendors used the advance in CPU and graphics performance to develop
more sophisticated user environments. Graphical User Interfaces, or GUIs,
became standard and enabled users to start and run multiple programs in
the same user environment. Networking on PCs became pervasive.
This rapid growth came at a price: increased user expectations. Users
expected to be able to send e-mail while listening to streaming audio that
was being delivered via an Internet radio station. Users expected their
computing platform to be quick and responsive. Users expected
applications to start quickly and handle inconvenient background tasks,
such as automatically saving a file with minimal disruption. These
challenges are the problems that face software developers today.

Motivation for Concurrency in Software

Most end users have a simplistic view of complex computer systems.
Consider the following scenario: A traveling businessman has just come
back to his hotel after a long day of presentations. Too exhausted to go out,
he decides to order room service and stay in his room to watch his favorite
baseball team play. Given that hes on the road, and doesnt have access to
the game on his TV, he decides to check out the broadcast via the Internet.
His view of the system might be similar to the one shown in Figure 1.1.

Figure 1.1 End User View of Streaming Multimedia Content via the Internet

The users expectations are based on conventional broadcast delivery

systems which provide continuous, uninterrupted delivery of content. The
user does not differentiate between streaming the content over the Internet
and delivering the data via a broadcast network. To the user, watching a
baseball game on a laptop seems like a simple, straightforward task.
 Chapter 1: Introduction to Multi-Core Architecture 3

The reality is that the implementation of such a system is far more

difficult. From the client side, the PC must be able to download the
streaming video data, decompress/decode it, and draw it on the video
display. In addition, it must handle any streaming audio that accompanies
the video stream and send it to the soundcard. Meanwhile, given the
general purpose nature of the computer, the operating system might be
configured to run a virus scan or some other system tasks periodically.
On the server side, the provider must be able to receive the original
broadcast, encode/compress it in near real-time, and then send it over
the network to potentially hundreds of thousands of clients. A system
designer who is looking to build a computer system capable of streaming
a Web broadcast might look at the system as its shown in Figure 1.2.

Figure 1.2 End-to-End Architecture View of Streaming Multimedia Content

over the Internet

Contrast this view of a streaming multimedia delivery service with

the end users perspective of the system shown in Figure 1.1. In order to
provide an acceptable end-user experience, system designers must be
able to effectively manage many independent subsystems that operate in
parallel.
Careful inspection of Figure 1.2 shows that the problem of streaming
media content may be broken into a number of disparate parts; each acting
4 Multi-Core Programming

1
independently from one another. This decomposition allows us to break
down each task into a single isolated problem, making the problem much
more manageable.
Concurrency in software is a way to manage the sharing of resources
used at the same time. Concurrency in software is important for several
reasons:
Concurrency allows for the most efficient use of system resources.
Efficient resource utilization is the key to maximizing perform-
ance of computing systems. Unnecessarily creating dependencies
on different components in the system drastically lowers overall
system performance. In the aforementioned streaming media example,
one might naively take this, serial, approach on the client side:
1. Wait for data to arrive on the network
2. Uncompress the data
3. Decode the data
4. Send the decoded data to the video/audio hardware
This approach is highly inefficient. The system is completely idle
while waiting for data to come in from the network. A better
approach would be to stage the work so that while the system is
waiting for the next video frame to come in from the network,
the previous frame is being decoded by the CPU, thereby improving
overall resource utilization.
Many software problems lend themselves to simple concurrent
implementations. Concurrency provides an abstraction for
implementing software algorithms or applications that are naturally
parallel. Consider the implementation of a simple FTP server.
Multiple clients may connect and request different files. A single-
threaded solution would require the application to keep track
of all the different state information for each connection. A
more intuitive implementation would create a separate thread for
each connection. The connection state would be managed by this
separate entity. This multi-threaded approach provides a solution
that is much simpler and easier to maintain.
Its worth noting here that the terms concurrent and parallel are not
interchangeable in the world of parallel programming. When multiple

1
The term independently is used loosely here. Later chapters discuss the managing of
interdependencies that is inherent in multi-threaded programming.
 Chapter 1: Introduction to Multi-Core Architecture 5

software threads of execution are running in parallel, it means that the

active threads are running simultaneously on different hardware
resources, or processing elements. Multiple threads may make progress
simultaneously. When multiple software threads of execution are
running concurrently, the execution of the threads is interleaved onto a
single hardware resource. The active threads are ready to execute, but
only one thread may make progress at a given point in time. In order to
have parallelism, you must have concurrency exploiting multiple
hardware resources.

Parallel Computing Platforms

In order to achieve parallel execution in software, hardware must
provide a platform that supports the simultaneous execution of multiple
threads. Generally speaking, computer architectures can be classified by
two different dimensions. The first dimension is the number of
instruction streams that a particular computer architecture may be able
to process at a single point in time. The second dimension is the number
of data streams that can be processed at a single point in time. In this
way, any given computing system can be described in terms of how
instructions and data are processed. This classification system is known
as Flynns taxonomy (Flynn, 1972), and is graphically depicted in
Figure 1.3.

Figure 1.3 Flynns Taxonomy

6 Multi-Core Programming

Flynns taxonomy places computing platforms in one of four

categories:
A single instruction, single data (SISD) machine is a traditional
sequential computer that provides no parallelism in hardware.
Instructions are executed in a serial fashion. Only one data stream
is processed by the CPU during a given clock cycle. Examples of
these platforms include older computers such as the original IBM
PC, older mainframe computers, or many of the 8-bit home
computers such as the Commodore 64 that were popular in the
early 1980s.
A multiple instruction, single data (MISD) machine is capable of
processing a single data stream using multiple instruction streams
simultaneously. In most cases, multiple instruction streams need
multiple data streams to be useful, so this class of parallel
computer is generally used more as a theoretical model than a
practical, mass-produced computing platform.
A single instruction, multiple data (SIMD) machine is one in
which a single instruction stream has the ability to process
multiple data streams simultaneously. These machines are useful
in applications such as general digital signal processing, image
processing, and multimedia applications such as audio and video.
Originally, supercomputers known as array processors or vector
processors such as the Cray-1 provided SIMD processing
capabilities. Almost all computers today implement some form of
SIMD instruction set. Intel processors implement the MMX,
Streaming SIMD Extensions (SSE), Streaming SIMD Extensions 2
(SSE2), and Streaming SIMD Extensions 3 (SSE3) instructions that
are capable of processing multiple data elements in a single clock.
The multiple data elements are stored in the floating point

registers. PowerPC processors have implemented the AltiVec
instruction set to provide SIMD support.
A multiple instruction, multiple data (MIMD) machine is capable
of is executing multiple instruction streams, while working on a
separate and independent data stream. This is the most common
parallel computing platform today. New multi-core platforms

such as the Intel Core Duo processor fall into this category.
Given that modern computing machines are either the SIMD or MIMD
machines, software developers have the ability to exploit data-level and
task level parallelism in software.
 Chapter 1: Introduction to Multi-Core Architecture 7

Parallel Computing in Microprocessors

In 1965, Gordon Moore observed that the number of transistors available
to semiconductor manufacturers would double approximately every 18
to 24 months. Now known as Moores law, this observation has guided
computer designers for the past 40 years. Many people mistakenly think
of Moores law as a predictor of CPU clock frequency, and its not really
hard to understand why. The most commonly used metric in measuring
computing performance is CPU clock frequency. Over the past 40 years,
CPU clock speed has tended to follow Moores law. Its an important
distinction to make, however, as taking this view of Moores law
imposes unnecessary limits on a silicon designer. While improving
straight-line instruction throughput and clock speeds are goals worth
striving for, computer architects can take advantage of these extra
transistors in less obvious ways.
For example, in an effort to make the most efficient use of processor
resources, computer architects have used instruction-level parallelization
techniques to improve processor performance. Instruction-level parallelism
(ILP), also known as dynamic, or out-of-order execution, gives the CPU the
ability to reorder instructions in an optimal way to eliminate pipeline stalls.
The goal of ILP is to increase the number of instructions that are executed
2
by the processor on a single clock cycle . In order for this technique to be
effective, multiple, independent instructions must execute. In the case of
in-order program execution, dependencies between instructions may limit
the number of instructions available for execution, reducing the amount of
parallel execution that may take place. An alternative approach that
attempts to keep the processors execution units full is to reorder the
instructions so that independent instructions execute simultaneously. In
this case, instructions are executed out of program order. This dynamic
instruction scheduling is done by the processor itself. You will learn much
more about these techniques in a later chapter, but for now what is
important to understand is that this parallelism occurs at the hardware level
and is transparent to the software developer.
As software has evolved, applications have become increasingly
capable of running multiple tasks simultaneously. Server applications
today often consist of multiple threads or processes. In order to support
this thread-level parallelism, several approaches, both in software and
hardware, have been adopted.

2
A processor that is capable of executing multiple instructions in a single clock cycle is known as a
super-scalar processor.
8 Multi-Core Programming

One approach to address the increasingly concurrent nature of modern

software involves using a preemptive, or time-sliced, multitasking operating
system. Time-slice multi-threading allows developers to hide latencies
associated with I/O by interleaving the execution of multiple threads. This
model does not allow for parallel execution. Only one instruction stream
can run on a processor at a single point in time.
Another approach to address thread-level parallelism is to increase
the number of physical processors in the computer. Multiprocessor
systems allow true parallel execution; multiple threads or processes run
simultaneously on multiple processors. The tradeoff made in this case is
increasing the overall system cost.
As computer architects looked at ways that processor architectures
could adapt to thread-level parallelism, they realized that in many cases,
the resources of a modern processor were underutilized. In order to
consider this solution, you must first more formally consider what a
thread of execution in a program is. A thread can be defined as a basic
unit of CPU utilization. It contains a program counter that points to the
current instruction in the stream. It contains CPU state information for
the current thread. It also contains other resources such as a stack.
A physical processor is made up of a number of different resources,
including the architecture statethe general purpose CPU registers and
interrupt controller registers, caches, buses, execution units, and branch
prediction logic. However, in order to define a thread, only the
architecture state is required. A logical processor can thus be created by
duplicating this architecture space. The execution resources are then
shared among the different logical processors. This technique is known
as simultaneous multi-threading, or SMT. Intels implementation of SMT
is known as Hyper-Threading Technology, or HT Technology. HT
Technology makes a single processor appear, from softwares
perspective, as multiple logical processors. This allows operating systems
and applications to schedule multiple threads to logical processors as
they would on multiprocessor systems. From a microarchitecture
perspective, instructions from logical processors are persistent and
execute simultaneously on shared execution resources. In other words,
multiple threads can be scheduled, but since the execution resources are
shared, its up to the microarchitecture to determine how and when to
interleave the execution of the two threads. When one thread stalls,
another thread is allowed to make progress. These stall events include
handling cache misses and branch mispredictions.
The next logical step from simultaneous multi-threading is the multi-core
processor. Multi-core processors use chip multiprocessing (CMP). Rather
 Chapter 1: Introduction to Multi-Core Architecture 9

than just reuse select processor resources in a single-core processor,

processor manufacturers take advantage of improvements in manufacturing
technology to implement two or more execution cores within a single
processor. These cores are essentially two individual processors on a single
die. Execution cores have their own set of execution and architectural
resources. Depending on design, these processors may or may not share a
large on-chip cache. In addition, these individual cores may be combined
with SMT; effectively increasing the number of logical processors by twice
the number of execution cores. The different processor architectures are
highlighted in Figure 1.4.

A) Single Core B) Multiprocessor

C) Hyper-Threading Technology D) Multi-core

E) Multi-core with Shared Cache

F) Multi-core with Hyper-Threading Technology

Figure 1.4 Simple Comparison of Single-core, Multi-processor, and Multi-Core

Architectures
10 Multi-Core Programming

Differentiating Multi-Core Architectures from Hyper-Threading Technology

With HT Technology, parts of the one processor are shared between threads,
while other parts are duplicated between them. One of the most important
shared resources is the actual execution engine. This engine works on both
threads at the same time by executing instructions for one thread on
resources that the other thread is not using. When both threads are running,
HT Technology literally interleaves the instructions in the execution pipeline.
Which instructions are inserted when depends wholly on what execution
resources of the processor are available at execution time. Moreover, if one
thread is tied up reading a large data file from disk or waiting for the user to
type on the keyboard, the other thread takes over all the processor
resourceswithout the operating system switching tasksuntil the first
thread is ready to resume processing. In this way, each thread receives the
maximum available resources and the processor is kept as busy as possible.
An example of a thread running on a HT Technology enabled CPU is shown
in Figure 1.5.

Figure 1.5 Two Threads Executing on a Processor with Hyper-Threading

Technology

HT Technology achieves performance gains through latency hiding.

Fundamentally, a single execution core is shared among multiple threads.
Therefore, thread execution is not parallel. As a result, performance results
vary based on application and hardware platform. With HT Technology, in
certain applications, it is possible to attain, on average, a 30-percent increase
in processor throughput. In other words, in certain cases, the
processor can perform 1.3 times the number of executed instructions
that it could if it were running only one thread. To see a performance
improvement, applications must make good use of threaded
programming models and of the capabilities of Hyper-Threading
Technology.
 Chapter 1: Introduction to Multi-Core Architecture 11

The performance benefits of HT Technology depend on how much

latency hiding can occur in your application. In some applications,
developers may have minimized or effectively eliminated memory
latencies through cache optimizations. In this case, optimizing for HT
Technology may not yield any performance gains.
On the other hand, multi-core processors embed two or more
independent execution cores into a single processor package. By providing
multiple execution cores, each sequence of instructions, or thread, has a
hardware execution environment entirely to itself. This enables each thread
run in a truly parallel manner. An example of two threads running on a
dual-core processor is shown in Figure 1.6. Compare this with the HT
Technology example provided in Figure 1.5, and note that a dual-core
processor provides true parallel execution of each thread.

Figure 1.6 Two Threads on a Dual-Core Processor with each Thread Running
Independently

It should be noted that HT Technology does not attempt to deliver

multi-core performance, which can theoretically be close to a 100-percent,
or 2x improvement in performance for a dual-core system. HT Technology
is more of a facility in which the programmer may be able to use idle CPU
resources in order to accomplish more work. When combined with multi-
core technology, HT Technology can provide powerful optimization
opportunities, increasing system throughput substantially.

Multi-threading on Single-Core versus Multi-Core Platforms

At this point, many readers may be asking themselves what all the
commotion is about. The concept of multiple threads in a single process
space has been around for decades. Most modern applications use
threads in one fashion or another today. As a result, many developers are
12 Multi-Core Programming

already familiar with the concept of threading, and have probably

worked on applications that have multiple threads. There are however,
certain important considerations developers should be aware of when
writing applications targeting multi-core processors:
Optimal application performance on multi-core architectures will
be achieved by effectively using threads to partition software
workloads. Many applications today use threads as a tool to improve
user responsiveness on single-core platforms. Rather than blocking
the user interface (UI) on a time consuming database query or disk
access, an application will spawn a thread to process the users
request. This allows the scheduler to individually schedule the main
control loop task that receives UI events as well as the data
processing task that is running the database query. In this model,
developers rely on straight-line instruction throughput improvements
to improve application performance.
This is the significant limitation of multi-threading on single-core
processors. Since single-core processors are really only able to
interleave instruction streams, but not execute them simultaneously,
the overall performance gains of a multi-threaded application on
single-core architectures are limited. On these platforms, threads are
generally seen as a useful programming abstraction for hiding latency.
This performance restriction is removed on multi-core architectures.
On multi-core platforms, threads do not have to wait for any one
resource. Instead, threads run independently on separate cores. As an
example, consider two threads that both wanted to execute a shift
operation. If a core only had one shifter unit they could not run in
parallel. On two cores, there would be two shifter units, and each
thread could run without contending for the same resource.
Multi-core platforms allow developers to optimize applications by
intelligently partitioning different workloads on different processor
cores. Application code can be optimized to use multiple processor
resources, resulting in faster application performance.
Multi-threaded applications running on multi-core platforms have
different design considerations than do multi-threaded applications
running on single-core platforms. On single-core platforms,
assumptions may be made by the developer to simplify writing and
debugging a multi-threaded application. These assumptions may not be
valid on multi-core platforms. Two areas that highlight these differences
are memory caching and thread priority.
 Chapter 1: Introduction to Multi-Core Architecture 13

In the case of memory caching, each processor core may have its
3
own cache. At any point in time, the cache on one processor core
may be out of sync with the cache on the other processor core. To
help illustrate the types of problems that may occur, consider the
following example. Assume two threads are running on a dual-core
processor. Thread 1 runs on core 1 and thread 2 runs on core 2. The
threads are reading and writing to neighboring memory locations.
Since cache memory works on the principle of locality, the data
values, while independent, may be stored in the same cache line. As a
result, the memory system may mark the cache line as invalid, even
though the data that the thread is interested in hasnt changed. This
problem is known as false sharing. On a single-core platform, there
is only one cache shared between threads; therefore, cache
synchronization is not an issue.
Thread priorities can also result in different behavior on single-core
versus multi-core platforms. For example, consider an application
that has two threads of differing priorities. In an attempt to improve
performance, the developer assumes that the higher priority thread
will always run without interference from the lower priority thread.
On a single-core platform, this may be valid, as the operating systems
scheduler will not yield the CPU to the lower priority thread.
However, on multi-core platforms, the scheduler may schedule both
threads on separate cores. Therefore, both threads may run
simultaneously. If the developer had optimized the code to assume
that the higher priority thread would always run without interference
from the lower priority thread, the code would be unstable on multi-
core and multi-processor systems.
One goal of this book is to help developers correctly utilize the number
of processor cores they have available.

Understanding Performance
At this point one may wonderhow do I measure the performance
benefit of parallel programming? Intuition tells us that if we can
subdivide disparate tasks and process them simultaneously, were likely

3
Multi-core CPU architectures can be designed in a variety of ways: some multi-core CPUs will share the
on-chip cache between execution units; some will provide a dedicated cache for each execution core;
and others will take a hybrid approach, where the cache is subdivided into layers that are dedicated to a
particular execution core and other layers that are shared by all execution cores. For the purposes of
this section, we assume a multi-core architecture with a dedicated cache for each core.
14 Multi-Core Programming

to see significant performance improvements. In the case where the tasks

are completely independent, the performance benefit is obvious, but
most cases are not so simple. How does one quantitatively determine the
performance benefit of parallel programming? One metric is to compare
the elapsed run time of the best sequential algorithm versus the elapsed
run time of the parallel program. This ratio is known as the speedup and
characterizes how much faster a program runs when parallelized.
Timebest _ sequential _ algorithm
Speedup(nt) =
Timeparallel _ implementation ( nt )

Speedup is defined in terms of the number of physical threads (nt)

used in the parallel implementation.

Amdahls Law
Given the previous definition of speedup, is there a way to determine the
theoretical limit on the performance benefit of increasing the number of
processor cores, and hence physical threads, in an application? When
examining this question, one generally starts with the work done by
Gene Amdahl in 1967. His rule, known as Amdahls Law, examines the
maximum theoretical performance benefit of a parallel solution relative
to the best case performance of a serial solution.
Amdahl started with the intuitively clear statement that program
speedup is a function of the fraction of a program that is accelerated and
by how much that fraction is accelerated.
1
Speedup =
(1 FractionEnhanced ) + (FractionEnhanced/SpeedupEnhanced )

So, if you could speed up half the program by 15 percent, youd get:
Speedup = 1 / ((1 .50) + (.50/1.15)) = 1 / (.50 + .43) = 1.08

This result is a speed increase of 8 percent, which is what youd expect.

If half of the program is improved 15 percent, then the whole program is
improved by half that amount.
Amdahl then went on to explain how this equation works out if you
make substitutions for fractions that are parallelized and those that are
run serially, as shown in Equation 1.1.
 Chapter 1: Introduction to Multi-Core Architecture 15

Equation 1.1 Amdahls Law

1
Speedup =
S + (1 S )/n

In this equation, S is the time spent executing the serial portion of the
parallelized version and n is the number of processor cores. Note that the
numerator in the equation assumes that the program takes 1 unit of time
to execute the best sequential algorithm.
If you substitute 1 for the number of processor cores, you see that no
speedup is realized. If you have a dual-core platform doing half the work,
the result is:
1 / (0.5S + 0.5S/2) = 1/0.75S = 1.33
or a 33-percent speed-up, because the run time, as given by the
denominator, is 75 percent of the original run time. For an 8-core
processor, the speedup is:
1 / (0.5S + 0.5S/8) = 1/0.75S = 1.78
Setting n = in Equation 1.1, and assuming that the best sequential
algorithm takes 1 unit of time yields Equation 1.2.

Equation 1.2 Upper Bound of an Application with S Time Spent in Sequential

Code

1
Speedup =
S

As stated in this manner, Amdahl assumes that the addition of processor

cores is perfectly scalable. As such, this statement of the law shows the
maximum benefit a program can expect from parallelizing some portion
of the code is limited by the serial portion of the code. For example,
according Amdahls law, if 10 percent of your application is spent in serial
code, the maximum speedup that can be obtained is 10x, regardless of the
number of processor cores.
It is important to note that endlessly increasing the processor cores only
affects the parallel portion of the denominator. So, if a program is only 10-
percent parallelized, the maximum theoretical benefit is that the program
can run in 90 percent of the sequential time.
16 Multi-Core Programming

Given this outcome, you can see the first corollary of Amdahls
law: decreasing the serialized portion by increasing the parallelized
portion is of greater importance than adding more processor cores. For
example, if you have a program that is 30-percent parallelized running on
a dual-core system, doubling the number of processor cores reduces run
time from 85 percent of the serial time to 77.5 percent, whereas
doubling the amount of parallelized code reduces run time from 85
percent to 70 percent. This is illustrated in Figure 1.7. Only when a
program is mostly parallelized does adding more processors help more
than parallelizing the remaining code. And, as you saw previously, you
have hard limits on how much code can be serialized and on how many
additional processor cores actually make a difference in performance.

Performance benefit of doubling Performance benefit of doubling

the number of processor cores the amount of parallelism in code
Note: The advantage gained by writing parallel code

Figure 1.7 Theoretical Performance Comparison between Increasing Number

of CPU Cores versus Increasing Concurrency in Implementation

To make Amdahls Law reflect the reality of multi-core systems, rather

than the theoretical maximum, system overhead from adding threads
should be included:
1
Speedup =
S + (1 S )/n + H ( n )

where H(n) = overhead, and again, we assume that the best serial
algorithm runs in one time unit. Note that this overhead is not linear on a
good parallel machine.
 Chapter 1: Introduction to Multi-Core Architecture 17

This overhead consists of two portions: the actual operating system

overhead and inter-thread activities, such as synchronization and other forms
of communication between threads. Notice that if the overhead is big
enough, it offsets the benefits of the parallelized portion. In fact, if the
overhead is large enough, the speedup ration can ultimately have a value of
less than 1, implying that threading has actually slowed performance when
compared to the single-threaded solution. This is very common in poorly
architected multi-threaded applications. The important implication is that
the overhead introduced by threading must be kept to a minimum. For this
reason, most of this book is dedicated to keeping the cost of threading as
low as possible.

Amdahls Law Applied to Hyper-Threading Technology

The previous section demonstrated Amdahls law as it applies to multi-
processor and multi-core systems. Hyper-Threading Technology
imposes an additional factor on how you apply Amdahls Law to your
code. On processors enabled with HT Technology, the fact that certain
processor resources are shared between the different threads of
execution has a direct effect on the maximum performance benefit of
threading an application.
Given the interleaved execution environment provided by HT
Technology, its important to develop a form of Amdahls law that works
for HT Technology. Assume that your application experiences a
performance gain of around 30 percent when run on a processor with
HT Technology enabled. That is, performance improves by 30 percent
over the time required for a single processor to run both threads. If you
were using a quad-core platform, with each processor completely
dedicated to the thread it was running, the number could, in theory,
be up to 4x. That is, the second, third, and fourth processor core
could give a 300-percent boost to program throughput. In practice its
not quite 300 percent, due to overhead and code that cannot be
parallelized, and the performance benefits will vary based on the
application.
Inside the processor enabled with HT Technology, each thread is
running more slowly than it would if it had the whole processor to itself.
HT Technology is not a replacement for multi-core processing since
many processing resources, such as the execution units, are shared. The
slowdown varies from application to application. As example, assume
each thread runs approximately one-third slower than it would if it
18 Multi-Core Programming

owned the entire processor. Amending Amdahls Law to fit HT

Technology, then, you get:
1
SpeedupHTT =
S + 0.67((1 S )/n ) + H ( n )

where n = number of logical processors.

This equation represents the typical speed-up for programs running
on processor cores with HT Technology performance. The value of H(n)
is determined empirically and varies from application to application.

Growing Returns: Gustafsons Law

Based on Amdahls work, the viability of massive parallelism was
questioned for a number of years. Then, in the late 1980s, at the Sandia
National Lab, impressive linear speedups in three practical applications
were observed on a 1,024-processor hypercube. The results (Gustafson
1988) demonstrated that near linear speedup was possible in many
practical cases, even when Amdahls Law predicted otherwise.
Built into Amdahls Law are several assumptions that may not hold true
in real-world implementations. First, Amdahls Law assumes that the best
performing serial algorithm is strictly limited by the availability of CPU
cycles. This may not be the case. A multi-core processor may implement a
separate cache on each core. Thus, more of the problems data set may be
stored in cache, reducing memory latency. The second flaw is that
Amdahls Law assumes that the serial algorithm is the best possible
solution for a given problem. However, some problems lend themselves to
a more efficient parallel solution. The number of computational steps may
be significantly less in the parallel implementation.
Perhaps the biggest weakness, however, is the assumption that
Amdahls Law makes about the problem size. Amdahls Law assumes that
as the number of processor cores increases, the problem size stays the
same. In most cases, this is not valid. Generally speaking, when given
more computing resources, the problem generally grows to meet the
resources available. In fact, it is more often the case that the run time of
the application is constant.
Based on the work at Sandia, an alternative formulation for speedup,
referred to as scaled speedup was developed by E. Barsis.
Scaled speedup = N + (1 N ) * s

where N = is the number of processor cores and s is the ratio of the time
spent in the serial port of the program versus the total execution time.
 Chapter 1: Introduction to Multi-Core Architecture 19

Scaled speedup is commonly referred to as Gustafsons Law. From this

equation, one can see that the speedup in this case is linear.
Gustafsons Law has been shown to be equivalent to Amdahls Law
(Shi 1996). However, Gustafsons Law offers a much more realistic look
at the potential of parallel computing on multi-core processors.

Key Points
This chapter demonstrated the inherent concurrent nature of many
software applications and introduced the basic need for parallelism in
hardware. An overview of the different techniques for achieving parallel
execution was discussed. Finally, the chapter examined techniques for
estimating the performance benefits of using proper multi-threading
techniques. The key points to keep in mind are:
Concurrency refers to the notion of multiple threads in progress
at the same time. This is often achieved on sequential processors
through interleaving.
Parallelism refers to the concept of multiple threads executing
simultaneously.
Modern software applications often consist of multiple processes
or threads that can be executed in parallel.
Most modern computing platforms are multiple instruction,
multiple data (MIMD) machines. These machines allow
programmers to process multiple instruction and data streams
simultaneously.
In practice, Amdahls Law does not accurately reflect the benefit
of increasing the number of processor cores on a given platform.
Linear speedup is achievable by expanding the problem size with
the number of processor cores.
 Chapter 2
System Overview
of Threading
W hen implemented properly, threading can enhance performance by
making better use of hardware resources. However, the improper
use of threading can lead to degraded performance, unpredictable
behavior, and error conditions that are difficult to resolve. Fortunately, if
you are equipped with a proper understanding of how threads operate,
you can avoid most problems and derive the full performance benefits
that threads offer. This chapter presents the concepts of threading
starting from hardware and works its way up through the operating
system and to the application level.
To understand threading for your application you need to understand
the following items:
The design approach and structure of your application
The threading application programming interface (API)
The compiler or runtime environment for your application
The target platforms on which your application will run
From these elements, a threading strategy can be formulated for use in
specific parts of your application.
Since the introduction of instruction-level parallelism, continuous
advances in the development of microprocessors have resulted in
processors with multiple cores. To take advantage of these multi-core
processors you must understand the details of the software threading
model as well as the capabilities of the platform hardware.

21
22 Multi-Core Programming

You might be concerned that threading is difficult and that you might
have to learn specialized concepts. While its true in general, in reality
threading can be simple, once you grasp the basic principles.

Defining Threads
A thread is a discrete sequence of related instructions that is executed
independently of other instruction sequences. Every program has at least
one threadthe main threadthat initializes the program and begins
executing the initial instructions. That thread can then create other
threads that perform various tasks, or it can create no new threads and
simply do all the work itself. In either case, every program has at least
one thread. Each thread maintains its current machine state.
At the hardware level, a thread is an execution path that remains
independent of other hardware thread execution paths. The operating
system maps software threads to hardware execution resources as
described later in this chapter
The decision to thread your application should reflect the needs of
the program and the basic execution capabilities of the deployment
platform. Not everything should be threaded. Too much threading can
hurt performance. As with many aspects of programming, thoughtful
design and proper testing determine the right balance.

System View of Threads

The thread computational model is represented in Figure 2.1. As
illustrated, there are three layers for threading:
User-level threads. Threads created and manipulated in the
application software.
Kernel-level threads. The way the operating system implements
most threads.
Hardware threads. How threads appear to the execution
resources in the hardware.
A single program thread frequently involves all three levels: a program
thread is implemented by the operating system as a kernel-level thread,
and executed as a hardware thread.
Between these layers are interfaces, which are frequently handled
automatically by the executing system. However, to make good use of
threading resources, its important to know how these interfaces work.
 Chapter 2: System Overview of Threading 23

They are touched on in this chapter and treated in greater detail in

Chapters 3, 4, and 5.

User-Level Threads

Operational Flow
Used by executable application and handled by user-level OS

Kernel-Level Threads
Used by operating system kernel and and
handled by kernal-level OS

Hardware Threads
Used by each Processor

Figure 2.1 Computation Model of Threading

Threading above the Operating System

Developers can best understand the problems they face using threads if
they know what actually takes place when threads are used in an
application. In applications that do not rely on a runtime framework, the
thread creation code is made as a call to system APIs. These calls are then
executed at runtime as calls to the operating system kernel to create a
thread. The instructions for the threads activity are then passed to the
processor for execution. Figure 2.2 shows the thread flow in a typical
system for traditional applications. In the Defining and Preparing stage,
threads are specified by the programming environment and encoded by
the compiler. During the Operating stage, threads are created and
managed by the operating system. Finally, in the Executing stage, the
processor executes the sequence of thread instructions.

Defining and Operating Executing

Preparing Threads Threads
Threads
Performed by Performed by OS Performed by Processors
Programming using Processes
Environment
and Compiler
Showing return trip to represent that after
execution operations get pass to user space

Figure 2.2 Flow of Threads in an Execution Environment

24 Multi-Core Programming

The application code might rely on a runtime environment. Such code,

known as managed code runs in an environment that performs some
application functions and makes calls the underlying operating system.
Managed environments include both Java Virtual Machine (JVM) and
Microsofts Common Language Runtime (CLR). These environments do not
provide any scheduling capability, relying instead on the operating system
for scheduling. Threading is passed to the operating system scheduler,
which handles the remaining downstream thread activity.
In general, application threads can be implemented at the application

level using established APIs. The most common APIs are OpenMP and
explicit low-level threading libraries such as Pthreads and Windows threads.
The choice of API depends on the requirements and the system platform. In
general, low-level threading requires significantly more code than solutions
such as OpenMP; the benefit they deliver, however, is fine-grained control
over the programs use of threads. OpenMP, in contrast, offers ease of use
and a more developer-friendly threading implementation. OpenMP requires
a compiler that supports the OpenMP API. Today, these are limited to
C/C++ and Fortran compilers. Coding low-level threads requires only access
to the operating systems multi-threading libraries. For further details on
OpenMP, Pthreads, and Windows threads, see Chapters 5 and 6.
To show how threading is used in a program, Listing 2.1 and Listing 2.2
are simple Hello World programs that use the OpenMP and Pthreads
libraries, respectively.

#include <stdio.h>
// Have to include 'omp.h' to get OpenMP definitons
#include <omp.h>
void main()
{
int threadID, totalThreads;
/* OpenMP pragma specifies that following block is
going to be parallel and the threadID variable is
private in this openmp block. */

#pragma omp parallel private(threadID)

{
threadID = omp_get_thread_num();
printf("\nHello World is from thread %d\n",
(int)threadID);
/* Master thread has threadID = 0 */
if (threadID == 0) {
printf("\nMaster thread being called\n");
totalThreads = omp_get_num_threads();
 Chapter 2: System Overview of Threading 25

printf("Total number of threads are %d\n",

totalThreads);
}
}
}

Listing 2.1 Hello World Program Using OpenMP

#include <pthread.h>
#include <stdio.h>
#include <stdlib.h>
#define NUM_THREADS 5

void *PrintHello(void *threadid)

{
printf("\n%d: Hello World!\n", threadid);
pthread_exit(NULL);
}

int main(int argc, char *argv[])

{
pthread_t threads[NUM_THREADS];
int rc, t;
for (t=0; t < NUM_THREADS; t++) {
printf("Creating thread %d\n", t);
rc = pthread_create( &threads[t], NULL,
PrintHello,(void *)t);
if (rc) {
printf("ERROR return code from pthread_create(): %d\n",
rc);
exit(-1);
}
}
pthread_exit(NULL);
}

Listing 2.2 Hello World Program Using Pthreads

As can be seen, the OpenMP code in Listing 2.1 has no function that
corresponds to thread creation. This is because OpenMP creates threads
automatically in the background. Explicit low-level coding of threads
is more evident in Pthreads, shown in Listing 2.2, where a call to
pthread_create() actually creates a single thread and points it at the
work to be done in PrintHello().
26 Multi-Core Programming

Threads inside the OS

The key to viewing threads from the perspective of a modern operating
system is to recognize that operating systems are partitioned into two
distinct layers: the user-level partition (where applications are run) and the
kernel-level partition (where system oriented activities occur). Figure 2.3
shows these partitions along with other components. This figure shows
the interface between application layer and the kernel-level operating
system, referred to as system libraries. These contain the necessary
operating-system components that can be run with user-level privilege. As
illustrated, the interface between the operating system and the processor is
the hardware abstraction layer (HAL).

Application Layer
Applications and Required Service Components
User-Level
OS

System Libraries
Kernel-Level

Process, Kernel
OS

Other
Threads and IO Memory Internal
Operational
Resource Manager Manager Operational
Units
Scheduler Manager

HAL
(Hardware Abstraction Layer)

Architecture
(Proccessors and Chipset)

Figure 2.3 Different Layers of the Operating System

The kernel is the nucleus of the operating system and maintains

tables to keep track of processes and threads. The vast majority of thread-
level activity relies on kernel-level threads. Threading libraries such as
OpenMP and Pthreads (POSIX standard threads) use kernel-level threads.
Windows supports both kernel-level and user-level threads. User-level
 Chapter 2: System Overview of Threading 27

threads, which are called fibers on the Windows platform, require the
programmer to create the entire management infrastructure for the
threads and to manually schedule their execution. Their benefit is that
the developer can manipulate certain details that are obscured in kernel-
level threads. However, because of this manual overhead and some
additional limitations, fibers might not add much value for well designed
multi-threaded applications.
Kernel-level threads provide better performance, and multiple kernel
threads from the same process can execute on different processors or
cores. The overhead associated with kernel-level threading is higher than
user-level threading and so kernel-level threads are frequently reused
once they have finished their original work.
Processes are discrete program tasks that have their own address space.
They are the coarse-level execution unit maintained as an independent entity
inside an operating system. There is a direct correlation between processes
and threads. Multiple threads can reside in a process. All threads in a process
share the same address space and so they benefit from simple inter-thread
communication. Instead of maintaining an individual process-based thread
list, the kernel maintains a thread table to keep track of all threads. The
operating system assigns a process control block (PCB) to each process; it
contains data on the processs unique identity, current machine state, the
priority of the process, and the address of the virtual memory where the
process resides.
Figure 2.4 shows the relationship between processors, processes, and
threads in modern operating systems. A processor runs threads from one
or more processes, each of which contains one or more threads.

Processors Processor

Processes Map to
OP1 OP2 OPn MMU

Map to
Threads T 1 T2 Tm Processors

: Processor OP1 : Process T1 : Thread MMU : Main Memory Unit

Figure 2.4 Relationships among Processors, Processes, and Threads

28 Multi-Core Programming

A program has one or more processes, each of which contains one or

more threads, each of which is mapped to a processor by the scheduler
in the operating system. A concept known as processor affinity enables
the programmer to request mapping of a specific thread to a specific
processor. Most operating systems today attempt to obey these requests,
but they do not guarantee fulfillment.
Various mapping models are used between threads and processors:
one to one (1:1), many to one (M:1), and many to many (M:N), as shown
in Figure 2.5. The 1:1 model requires no thread-library scheduler
overhead and the operating system handles the thread scheduling
responsibility. This is also referred to as preemptive multi-threading.
Linux, Windows 2000, and Windows XP use this preemptive multi-
threading model. In the M:1 model, the library scheduler decides which
thread gets the priority. This is called cooperative multi-threading. In
the case of M:N, the mapping is flexible.

Use-level
threads

User Space
Processes

Operating
System Threads
Kernel Space

Operating
System
Scheduler
HAL
Hardware

P/C P/C P/C P/C

TLS : Thread Level Scheduler HAL : Hardware Abstraction Layer

P/C : Processor or Core

(a) 1:1 Mapping of Threads to Processors

Figure 2.5 Mapping Models of Threads to Processors

 Chapter 2: System Overview of Threading 29

Use-level
threads

User Space
Processes

TLS TLS TLS TLS TLS

Operating
System Threads

Kernel Space
Operating
System
Scheduler
HAL

Hardware
P/C P/C P/C P/C

(b) M:1 Mapping of Threads to Processors

Use-level
threads

User Space
Processes

TLS TLS TLS TLS TLS

Operating

Kernel Space
System Threads

Operating
System
HAL
Scheduler

Hardware
P/C P/C P/C P/C

(c) M:N Mapping of Threads to Processors

Figure 2.5 Mapping Models of Threads to Processors (continued)

User-level threads such as those in Windows are mapped to kernel

threads; and so, when they are executing, the processor knows them
only as kernel-level threads.
In general, a preemptive or 1:1 model enables stronger handling of
the threads by the operating system. This book focuses only on Windows
and Linux and so it emphasizes this mode. For other operating systems,
see the References section.
30 Multi-Core Programming

Threads inside the Hardware

The hardware executes the instructions from the software levels.
Instructions of your application threads are mapped to resources and
flow down through the intermediate componentsthe operating system,
runtime environment, and virtual layerto the hardware.
Threading on hardware once required multiple CPUs to implement
parallelism: each thread ran on its own separate processor. Today,
processors with Hyper-Threading Technology (HT Technology) and
multiple cores provide multi-threading on a single processor. These
multi-threaded processors allow two or more threads of execution to run
on a single CPU at the same time. This CPU might have only one
execution engine or core but share the pipeline and other hardware
resources among the executing threads. Such processing would be
considered concurrent but not parallel; Figure 2.6 illustrates this
difference.

Operational Operational
Path Path
Operational Time

Concurrency Parallelism

Figure 2.6 Concurrency versus Parallelism

Multi-core CPUs, however, provide two or more execution cores, and

so they deliver true hardware-based multi-threading. Because both
threads execute on the same processor, this design is sometimes referred
 Chapter 2: System Overview of Threading 31

to as chip multi-threading (CMT). By contrast, HT Technology uses a

single core in which two threads share most of the execution resources.
This approach is called simultaneous multi-threading (SMT). SMT uses a
hardware scheduler to manage different hardware threads that are in
need of resources. The number of hardware threads that can execute
simultaneously is an important consideration in the design of software; to
achieve true parallelism, the number of active program threads should
always equal the number of available hardware threads. In most cases,
program threads will exceed the available hardware threads. However,
too many software threads can slow performance. So, keeping a balance
of software and hardware threads delivers good results.

What Happens When a Thread Is Created

As discussed earlier, there can be more than one thread in a process; and
each of those threads operates independently, even though they share
the same address space and certain resources, such as file descriptors.
In addition, each thread needs to have its own stack space. These stacks
are usually managed by the operating system. Figure 2.7 shows a typical
stack representation of a multi-threaded process. As an application
developer, you should not have to worry about the details of stack
management, such as thread stack sizes or thread stack allocation. On the
other hand, system-level developers must understand the underlying
details. If you want to use threading in your application, you must be
aware of the operating systems limits. For some applications, these
limitations might be restrictive, and in other cases, you might have to
bypass the default stack manager and manage stacks on your own. The
default stack size for a thread varies from system to system. That is why
creating many threads on some systems can slow performance
dramatically.
32 Multi-Core Programming

Address N
Stack

Stack
Region for Thread 1

Stack
Region for Thread 2

Heap

Program Code + Data

Address 0

Figure 2.7 Stack Layout in a Multi-threaded Process

Once created, a thread is always in one of four states: ready, running,

waiting (blocked), or terminated. There are additional sub-states that
reflect various reasons for entering one of the four basic states. These
finer sub-states can be valuable in debugging or analyzing a threaded
application.
Every process has at least one thread. This initial thread is created as
part of the process initialization. Application threads you create will run
while the initial thread continues to execute. As indicated in the state
diagram in Figure 2.8, each thread you create starts in a ready state.
Afterwards, when the new thread is attempting to execute instructions, it
is either in the running state or blocked. It is blocked if it is waiting for a
resource or for another thread. When a thread has completed its work, it
is either terminated or put back by the program into the ready state. At
program termination, the main thread and subsidiary threads are
terminated.
 Chapter 2: System Overview of Threading 33

Enter Exit
New Interrupt Terminate

Ready Running

Scheduler Dispatch
Event Compleation Event Wait
Stack
Waiting

Figure 2.8 State Diagram for a Thread

Application Programming Models and Threading

Threads are used liberally by the operating system for its own internal
activities so even if you write a single-threaded application, your runtime
setup will be heavily threaded. All major programming languages today
support the use of threads, whether those languages are imperative (C,
Fortran, Pascal, Ada), object-oriented (C++, Java, C#), functional (Lisp,
Miranda, SML), or logical (Prolog).

Virtual Environment: VMs and Platforms

One of the most important trends in computing today is virtualization.
Virtualization is the process of using computing resources to create the
appearance of a different set of resources. Runtime virtualization, such as
found in the Java JVM, creates the appearance to a Java application that it
is running in its own private environment or machine. System
virtualization creates the appearance of a different kind of virtual
machine, in which there exists a complete and independent instance of
the operating system. Both forms of virtual environments make effective
use of threads internally.
34 Multi-Core Programming

Runtime Virtualization
The operation of runtime virtualization is being provided by runtime
virtual machine. These virtual machines (VMs) can be considered as a
container and executor application on top of an operating system. There
are two mainstream VMs in use today: the Java VM and Microsofts
Common Language Runtime (CLR) that were discussed previously. These
VMs, for example, create at least three threads: the executing thread, a
garbage-collection thread that frees memory blocks that are no longer in
use, and a thread for just-in-time (JIT) compilation of bytecodes into
executable binary code. The VMs generally create other threads for
internal tasks. The VM and the operating system work in tandem to map
these threads to the available execution resources in a way that will
benefit performance as much as possible.

System Virtualization
System virtualization creates a different type of virtual machine. These
VMs recreate a complete execution context for software: they use
virtualized network adapters and disks and run their own instance of
the operating system. Several such VMs can run on the same hardware
platform, each with its separate operating system. The virtualization
layer that sits between the host system and these VMs is called the
virtual machine monitor (VMM). The VMM is also known as the
hypervisor. Figure 2.9 compares systems running a VMM with one that
does not.

App App App VM0 VM1

App App App App App App
Operating System ... ...
Device Guest OS0 Guest OS1
IDE NC . . . Drivers

Physical Host Hardware

Processors Memory Graphics VM Monitor (VMM)

Physical Host Hardware

Network Storage Keyboard/Mouse

Without VMs: Single OS owns With VMs: Multiple OSes share

all hardware resources hardware resources

Figure 2.9 Comparison of Systems without and with a VMM

 Chapter 2: System Overview of Threading 35

A VMM delivers the necessary virtualization of the underlying

platform such that the operating system in each VM runs under the
illusion that it owns the entire hardware platform.
When virtualizing the underlying hardware, VM software makes use
of a concept called virtual processors. It presents as many virtual
processors to the guest operating system as there are cores on the actual
host hardware. HT Technology does not change the number of virtual
processors, only cores count. One of the important benefits of processor
virtualization is that it can create isolation of the instruction-set
architecture (ISA). Certain processor instructions can be executed only
by the operating system because they are privileged instructions. On
todays Intel processors, only one piece of softwarethe host operating
systemhas this level of privilege. The VMM and the entire VM run as
applications. So, what happens when one of the guest operating systems
needs to run a privileged instruction? This instruction is trapped by the
virtual processor in the VM and a call is made to the VMM. In some cases,
the VMM can handle the call itself, in others it must pass the call on to
the host operating system, wait for the response and emulate that
response in the virtual processor. By this means, the VMM manages to
sidestep the execution of privileged instructions.
However, this process has a distinct performance cost associated
with it. As a result, Intel has developed a series of extensions to the ISA
that provide efficient ways for VMMs to execute the privileged
instructions of guest operating systems. These extensions are part of

Intel Virtualization Technology, and are designed to improve the
performance of VMMs.

Mapping Application Threads

VMMs do very little unusual to handle application threads. When an
application running in a VM creates a thread, the thread creation and
subsequent scheduling is all handled by the guest operating system. The
VMM does not need to know about it. When the guest operating system
schedules the thread, the virtual processor executes the instructions
using the same methods it executes any other sequence instructions. The
VMM makes no attempt to match application threads to specific
processor cores or to second-guess the guest operating systems
scheduler. So, on a system with a dual-core processor, for example, the
VMM presents two virtual processors to the guest operating system. That
OS then schedules threads on those processors as it would if it were
36 Multi-Core Programming

running on the actual hardware. The VMM executes the instructions but
pays little notice to what application threads are running.
The only time the VMM interrupts this process is when it needs to
swap out a VM or perform internal tasks. In such a case, several issues
can arise. For example, when a VMM is running multiple guest VMs, it
has to time-slice between them. Suppose a thread is locked and waiting
for a thread running on a different virtual processor when that other
processor is swapped out. The original VM will encounter a substantial
delay that would not occur if both VMs had been running on their own
dedicated hardware systems. This problem, known as lock-holder pre-
emption, is one of several that arise from the fact that guest VM
resources must be swapped out at times and the exact state of all threads
might not expect this situation. However, as virtualization becomes more
widely adopted, its likely that operating systems will offer features that
assist VMMs to coordinate this kind of activity.

Key Points
The concepts of threading depend on an understanding of the interaction
of various system components.
To properly comprehend the impact of threading, it is important
to understand the impact of threads on system components.
Software threads are different than hardware threads, but maintain
a direct relationship.
Application threading can be implemented using APIs or multi-
threading libraries.
Processes, threads, and fibers are different levels of the execution
mechanism within a system.
The thread life cycle has four stages: ready, running, waiting
(blocked), and terminated.
There are two types of virtualization on a system: runtime
virtualization and system virtualization.
A virtual machine monitor (VMM) typically makes no attempt to
match application threads to specific processor cores or to
second-guess the guest operating systems scheduler.
 Chapter 3
Fundamental
Concepts of Parallel
Programming
A s discussed in previous chapters, parallel programming uses threads
to enable multiple operations to proceed simultaneously. The entire
concept of parallel programming centers on the design, development,
and deployment of threads within an application and the coordination
between threads and their respective operations. This chapter examines
how to break up programming tasks into chunks that are suitable for
threading. It then applies these techniques to the apparently serial
problem of error diffusion.

Designing for Threads

Developers who are unacquainted with parallel programming generally
feel comfortable with traditional programming models, such as object-
oriented programming (OOP). In this case, a program begins at a defined
point, such as the main() function, and works through a series of tasks
in succession. If the program relies on user interaction, the main
processing instrument is a loop in which user events are handled. From
each allowed eventa button click, for example, the program performs
an established sequence of actions that ultimately ends with a wait for
the next user action.
When designing such programs, developers enjoy a relatively simple
programming world because only one thing is happening at any given
moment. If program tasks must be scheduled in a specific way, its
because the developer imposes a certain order on the activities. At any

37
38 Multi-Core Programming

point in the process, one step generally flows into the next, leading up to
a predictable conclusion, based on predetermined parameters.
To move from this linear model to a parallel programming model,
designers must rethink the idea of process flow. Rather than being
constrained by a sequential execution sequence, programmers should
identify those activities that can be executed in parallel. To do so, they
must see their programs as a set of tasks with dependencies between
them. Breaking programs down into these individual tasks and identifying
dependencies is known as decomposition. A problem may be decomposed
in several ways: by task, by data, or by data flow. Table 3.1 summarizes
these forms of decomposition. As you shall see shortly, these different
forms of decomposition mirror different types of programming activities.
Table 3.1 Summary of the Major Forms of Decomposition
Decomposition Design Comments
Task Different activities assigned to Common in GUI apps
different threads
Data Multiple threads performing the Common in audio
same operation but on different processing, imaging, and
blocks of data in scientific programming
Data Flow One threads output is the input Special care is needed to
to a second thread eliminate startup and
shutdown latencies

Task Decomposition
Decomposing a program by the functions that it performs is called task
decomposition. It is one of the simplest ways to achieve parallel
execution. Using this approach, individual tasks are catalogued. If two of
them can run concurrently, they are scheduled to do so by the
developer. Running tasks in parallel this way usually requires slight
modifications to the individual functions to avoid conflicts and to
indicate that these tasks are no longer sequential.
If we were discussing gardening, task decomposition would suggest
that gardeners be assigned tasks based on the nature of the activity: if
two gardeners arrived at a clients home, one might mow the lawn while
the other weeded. Mowing and weeding are separate functions broken
out as such. To accomplish them, the gardeners would make sure to have
some coordination between them, so that the weeder is not sitting in the
middle of a lawn that needs to be mowed.
 Chapter 3: Fundamental Concepts of Parallel Programming 39

In programming terms, a good example of task decomposition is

word processing software, such as Microsoft Word . When the user
opens a very long document, he or she can begin entering text right
away. While the user enters text, document pagination occurs in the
background, as one can readily see by the quickly increasing page count
that appears in the status bar. Text entry and pagination are two separate
tasks that its programmers broke out by function to run in parallel. Had
programmers not designed it this way, the user would be obliged to wait
for the entire document to be paginated before being able to enter any
text. Many of you probably recall that this wait was common on early
PC word processors.
Data Decomposition
Data decomposition, also known as data-level parallelism, breaks down
tasks by the data they work on rather than by the nature of the task.
Programs that are broken down via data decomposition generally have
many threads performing the same work, just on different data items. For
example, consider recalculating the values in a large spreadsheet. Rather
than have one thread perform all the calculations, data decomposition
would suggest having two threads, each performing half the calculations, or
th
n threads performing 1/n the work.
If the gardeners used the principle of data decomposition to divide
their work, they would both mow half the property and then both weed
half the flower beds. As in computing, determining which form of
decomposition is more effective depends a lot on the constraints of the
system. For example, if the area to mow is so small that it does not need
two mowers, that task would be better done by just one gardenerthat
is, task decomposition is the best choiceand data decomposition could
be applied to other task sequences, such as when the mowing is done
and both gardeners begin weeding in parallel.
As the number of processor cores increases, data decomposition
allows the problem size to be increased. This allows for more work to
be done in the same amount of time. To illustrate, consider the
gardening example. Two more gardeners are added to the work crew.
Rather than assigning all four gardeners to one yard, we can we can
assign the two new gardeners to another yard, effectively increasing our
total problem size. Assuming that the two new gardeners can perform
the same amount of work as the original two, and that the two yard
sizes are the same, weve doubled the amount of work done in the same
amount of time.
40 Multi-Core Programming

Data Flow Decomposition

Many times, when decomposing a problem, the critical issue isnt what
tasks should do the work, but how the data flows between the different
tasks. In these cases, data flow decomposition breaks up a problem by
how data flows between tasks.
The producer/consumer problem is a well known example of how data
flow impacts a programs ability to execute in parallel. Here, the output of
one task, the producer, becomes the input to another, the consumer. The
two tasks are performed by different threads, and the second one, the
consumer, cannot start until the producer finishes some portion of its work.
Using the gardening example, one gardener prepares the toolsthat
is, he puts gas in the mower, cleans the shears, and other similar tasks
for both gardeners to use. No gardening can occur until this step is
mostly finished, at which point the true gardening work can begin. The
delay caused by the first task creates a pause for the second task, after
which both tasks can continue in parallel. In computer terms, this
particular model occurs frequently.
In common programming tasks, the producer/consumer problem
occurs in several typical scenarios. For example, programs that must rely
on the reading of a file fit this scenario: the results of the file I/O become
the input to the next step, which might be threaded. However, that step
cannot begin until the reading is either complete or has progressed
sufficiently for other processing to kick off. Another common
programming example is parsing: an input file must be parsed, or
analyzed semantically, before the back-end activities, such as code
generation in a compiler, can begin.
The producer/consumer problem has several interesting dimensions:
The dependence created between consumer and producer can
cause significant delays if this model is not implemented
correctly. A performance-sensitive design seeks to understand the
exact nature of the dependence and diminish the delay it
imposes. It also aims to avoid situations in which consumer
threads are idling while waiting for producer threads.
In the ideal scenario, the hand-off between producer and
consumer is completely clean, as in the example of the file
parser. The output is context-independent and the consumer has
no need to know anything about the producer. Many times,
however, the producer and consumer components do not enjoy
 Chapter 3: Fundamental Concepts of Parallel Programming 41

such a clean division of labor, and scheduling their interaction

requires careful planning.
If the consumer is finishing up while the producer is completely
done, one thread remains idle while other threads are busy
working away. This issue violates an important objective of
parallel processing, which is to balance loads so that all available
threads are kept busy. Because of the logical relationship
between these threads, it can be very difficult to keep threads
equally occupied.
In the next section, well take a look at the pipeline pattern that allows
developers to solve the producer/consumer problem in a scalable
fashion.

Implications of Different Decompositions

Different decompositions provide different benefits. If the goal, for
example, is ease of programming and tasks can be neatly partitioned by
functionality, then task decomposition is more often than not the winner.
Data decomposition adds some additional code-level complexity to tasks, so
it is reserved for cases where the data is easily divided and performance is
important.
The most common reason for threading an application is
performance. And in this case, the choice of decompositions is more
difficult. In many instances, the choice is dictated by the problem
domain: some tasks are much better suited to one type of decomposition.
But some tasks have no clear bias. Consider for example, processing
images in a video stream. In formats with no dependency between
frames, youll have a choice of decompositions. Should they choose task
decomposition, in which one thread does decoding, another color
balancing, and so on, or data decomposition, in which each thread does
all the work on one frame and then moves on to the next? To return to
the analogy of the gardeners, the decision would take this form: If two
gardeners need to mow two lawns and weed two flower beds, how
should they proceed? Should one gardener only mowthat is, they
choose task based decompositionor should both gardeners mow
together then weed together?
In some cases, the answer emerges quicklyfor instance when a
resource constraint exists, such as only one mower. In others where
each gardener has a mower, the answer comes only through careful
analysis of the constituent activities. In the case of the gardeners, task
42 Multi-Core Programming

decomposition looks better because the start-up time for mowing is

saved if only one mower is in use. Ultimately, you determine the right
answer for your applications use of parallel programming by careful
planning and testing. The empirical timing and evaluation plays a more
significant role in the design choices you make in parallel programming
than it does in standard single-threaded programming.

Challenges Youll Face

The use of threads enables you to improve performance significantly by
allowing two or more activities to occur simultaneously. However,
developers cannot fail to recognize that threads add a measure of
complexity that requires thoughtful consideration to navigate correctly.
This complexity arises from the inherent fact that more than one activity
is occurring in the program. Managing simultaneous activities and their
possible interaction leads you to confronting four types of problems:
Synchronization is the process by which two or more threads
coordinate their activities. For example, one thread waits for
another to finish a task before continuing.
Communication refers to the bandwidth and latency issues
associated with exchanging data between threads.
Load balancing refers to the distribution of work across multiple
threads so that they all perform roughly the same amount of
work.
Scalability is the challenge of making efficient use of a larger
number of threads when software is run on more-capable
systems. For example, if a program is written to make good use of
four processor cores, will it scale properly when run on a system
with eight processor cores?
Each of these issues must be handled carefully to maximize application
performance. Subsequent chapters describe many aspects of these
problems and how best to address them on multi-core systems.

Parallel Programming Patterns

For years object-oriented programmers have been using design patterns
to logically design their applications. Parallel programming is no different
than object-oriented programmingparallel programming problems
 Chapter 3: Fundamental Concepts of Parallel Programming 43

generally fall into one of several well known patterns. A few of the more
common parallel programming patterns and their relationship to the
aforementioned decompositions are shown in Table 3.2.
Table 3.2 Common Parallel Programming Patterns
Pattern Decomposition
Task-level parallelism Task
Divide and Conquer Task/Data
Geometric Decomposition Data
Pipeline Data Flow
Wavefront Data Flow

In this section, well provide a brief overview of each pattern and the
types of problems that each pattern may be applied to.
Task-level Parallelism Pattern. In many cases, the best way to
achieve parallel execution is to focus directly on the tasks
themselves. In this case, the task-level parallelism pattern makes
the most sense. In this pattern, the problem is decomposed into a
set of tasks that operate independently. It is often necessary
remove dependencies between tasks or separate dependencies
using replication. Problems that fit into this pattern include the
so-called embarrassingly parallel problems, those where there
are no dependencies between threads, and replicated data
problems, those where the dependencies between threads may
be removed from the individual threads.
Divide and Conquer Pattern. In the divide and conquer pattern,
the problem is divided into a number of parallel sub-problems.
Each sub-problem is solved independently. Once each sub-
problem is solved, the results are aggregated into the final
solution. Since each sub-problem can be independently solved,
these sub-problems may be executed in a parallel fashion.
The divide and conquer approach is widely used on sequential
algorithms such as merge sort. These algorithms are very easy to
parallelize. This pattern typically does a good job of load
balancing and exhibits good locality; which is important for
effective cache usage.
Geometric Decomposition Pattern. The geometric decomposi-
tion pattern is based on the parallelization of the data structures
44 Multi-Core Programming

used in the problem being solved. In geometric decomposition,

each thread is responsible for operating on data chunks. This
pattern may be applied to problems such as heat flow and wave
propagation.
Pipeline Pattern. The idea behind the pipeline pattern is identical
to that of an assembly line. The way to find concurrency here is
to break down the computation into a series of stages and have
each thread work on a different stage simultaneously.
Wavefront Pattern. The wavefront pattern is useful when
processing data elements along a diagonal in a two-dimensional
grid. This is shown in Figure 3.1

Figure 3.1 Wavefront Data Access Pattern

The numbers in Figure 3.1 illustrate the order in which the data
elements are processed. For example, elements in the diagonal
that contains the number 3 are dependent on data elements
1 and 2 being processed previously. The shaded data
elements in Figure 3.1 indicate data that has already been
processed. In this pattern, it is critical to minimize the idle time
spent by each thread. Load balancing is the key to success with
this pattern.
For a more extensive and thorough look at parallel programming design
patterns, refer to the book Patterns for Parallel Programming (Mattson
2004).
 Chapter 3: Fundamental Concepts of Parallel Programming 45

A Motivating Problem: Error Diffusion

To see how you might apply the aforementioned methods to a practical
computing problem, consider the error diffusion algorithm that is used in
many computer graphics and image processing programs. Originally
proposed by Floyd and Steinberg (Floyd 1975), error diffusion is a
technique for displaying continuous-tone digital images on devices that
have limited color (tone) range. Printing an 8-bit grayscale image to a
black-and-white printer is problematic. The printer, being a bi-level
device, cannot print the 8-bit image natively. It must simulate multiple
shades of gray by using an approximation technique. An example of an
image before and after the error diffusion process is shown in Figure 3.2.
The original image, composed of 8-bit grayscale pixels, is shown on the
left, and the result of the image that has been processed using the error
diffusion algorithm is shown on the right. The output image is composed
of pixels of only two colors: black and white.

Original 8-bit image on the left, resultant 2-bit image on the right. At the resolution
of this printing, they look similar.

The same images as above but zoomed to 400 percent and cropped to 25 percent
to show pixel detail. Now you can clearly see the 2-bit black-white rendering on the
right and 8-bit gray-scale on the left.

Figure 3.2 Error Diffusion Algorithm Output

46 Multi-Core Programming

The basic error diffusion algorithm does its work in a simple three-
step process:
1. Determine the output value given the input value of the current
pixel. This step often uses quantization, or in the binary case,
thresholding. For an 8-bit grayscale image that is displayed on a 1-bit
output device, all input values in the range [0, 127] are to be
displayed as a 0 and all input values between [128, 255] are to
be displayed as a 1 on the output device.
2. Once the output value is determined, the code computes the
error between what should be displayed on the output device
and what is actually displayed. As an example, assume that the
current input pixel value is 168. Given that it is greater than our
threshold value (128), we determine that the output value will be
a 1. This value is stored in the output array. To compute the
error, the program must normalize output first, so it is in the
same scale as the input value. That is, for the purposes of
computing the display error, the output pixel must be 0 if the
output pixel is 0 or 255 if the output pixel is 1. In this case, the
display error is the difference between the actual value that
should have been displayed (168) and the output value (255),
which is 87.
3. Finally, the error value is distributed on a fractional basis to the
neighboring pixels in the region, as shown in Figure 3.3.

Figure 3.3 Distributing Error Values to Neighboring Pixels

This example uses the Floyd-Steinberg error weights to propagate

errors to neighboring pixels. 7/16ths of the error is computed and added
 Chapter 3: Fundamental Concepts of Parallel Programming 47

to the pixel to the right of the current pixel that is being processed.
5/16ths of the error is added to the pixel in the next row, directly below
the current pixel. The remaining errors propagate in a similar fashion.
While you can use other error weighting schemes, all error diffusion
algorithms follow this general method.
The three-step process is applied to all pixels in the image. Listing 3.1
shows a simple C implementation of the error diffusion algorithm, using
Floyd-Steinberg error weights.

/**************************************
* Initial implementation of the error diffusion algorithm.
***************************************/
void error_diffusion(unsigned int width,
unsigned int height,
unsigned short **InputImage,
unsigned short **OutputImage)
{
for (unsigned int i = 0; i < height; i++)
{
for (unsigned int j = 0; j < width; j++)
{
/* 1. Compute the value of the output pixel*/
if (InputImage[i][j] < 128)
OutputImage[i][j] = 0;
else
OutputImage[i][j] = 1;

/* 2. Compute the error value */

int err = InputImage[i][j] - 255*OutputImage[i][j];

/* 3. Distribute the error */

InputImage[i][j+1] += err * 7/16;
InputImage[i+1][j-1] += err * 3/16;
InputImage[i+1][j] += err * 5/16;
InputImage[i+1][j+1] += err * 1/16;
}
}
}

Listing 3.1 C-language Implementation of the Error Diffusion Algorithm

48 Multi-Core Programming

Analysis of the Error Diffusion Algorithm

At first glance, one might think that the error diffusion algorithm is an
inherently serial process. The conventional approach distributes errors to
neighboring pixels as they are computed. As a result, the previous pixels
error must be known in order to compute the value of the next pixel.
This interdependency implies that the code can only process one pixel at
a time. Its not that difficult, however, to approach this problem in a way
that is more suitable to a multithreaded approach.

An Alternate Approach: Parallel Error Diffusion

To transform the conventional error diffusion algorithm into an approach
that is more conducive to a parallel solution, consider the different
decomposition that were covered previously in this chapter. Which
would be appropriate in this case? As a hint, consider Figure 3.4, which
revisits the error distribution illustrated in Figure 3.3, from a slightly
different perspective.

Figure 3.4 Error-Diffusion Error Computation from the Receiving Pixels

Perspective

Given that a pixel may not be processed until its spatial predecessors
have been processed, the problem appears to lend itself to an approach
where we have a produceror in this case, multiple producers
producing data (error values) which a consumer (the current pixel) will
use to compute the proper output pixel. The flow of error data to the
current pixel is critical. Therefore, the problem seems to break down
into a data-flow decomposition.
 Chapter 3: Fundamental Concepts of Parallel Programming 49

Now that we identified the approach, the next step is to determine the
best pattern that can be applied to this particular problem. Each
independent thread of execution should process an equal amount of work
(load balancing). How should the work be partitioned? One way, based on
the algorithm presented in the previous section, would be to have a thread
that processed the even pixels in a given row, and another thread that
processed the odd pixels in the same row. This approach is ineffective
however; each thread will be blocked waiting for the other to complete,
and the performance could be worse than in the sequential case.
To effectively subdivide the work among threads, we need a way to
reduce (or ideally eliminate) the dependency between pixels. Figure 3.4
illustrates an important point that's not obvious in Figure 3.3that in
order for a pixel to be able to be processed, it must have three error
1
values (labeled eA, eB, and eC in Figure 3.3) from the previous row, and
one error value from the pixel immediately to the left on the current
row. Thus, once these pixels are processed, the current pixel may
complete its processing. This ordering suggests an implementation
where each thread processes a row of data. Once a row has completed
processing of the first few pixels, the thread responsible for the next row
may begin its processing. Figure 3.5 shows this sequence.

Multiple threads are able to process multiple rows simultaneously.

Figure 3.5 Parallel Error Diffusion for Multi-thread, Multi-row Situation

1
We assume eA = eD = 0 at the left edge of the page (for pixels in column 0); and that eC = 0 at the
right edge of the page (for pixels in column W-1, where W = the number of pixels in the image).
50 Multi-Core Programming

Notice that a small latency occurs at the start of each row. This
latency is due to the fact that the previous rows error data must be
calculated before the current row can be processed. These types of
latency are generally unavoidable in producer-consumer implementations;
however, you can minimize the impact of the latency as illustrated here.
The trick is to derive the proper workload partitioning so that each
thread of execution works as efficiently as possible. In this case, you
incur a two-pixel latency before processing of the next thread can begin.
An 8.5" X 11" page, assuming 1,200 dots per inch (dpi), would have
10,200 pixels per row. The two-pixel latency is insignificant here.
The sequence in Figure 3.5 illustrates the data flow common to the
wavefront pattern.

Other Alternatives
In the previous section, we proposed a method of error diffusion where
each thread processed a row of data at a time. However, one might
consider subdividing the work at a higher level of granularity.
Instinctively, when partitioning work between threads, one tends to look
for independent tasks. The simplest way of parallelizing this problem
would be to process each page separately. Generally speaking, each page
would be an independent data set, and thus, it would not have any
interdependencies. So why did we propose a row-based solution instead
of processing individual pages? The three key reasons are:
An image may span multiple pages. This implementation would
impose a restriction of one image per page, which might or might
not be suitable for the given application.
Increased memory usage. An 8.5 x 11-inch page at 1,200 dpi
consumes 131 megabytes of RAM. Intermediate results must be
saved; therefore, this approach would be less memory efficient.
An application might, in a common use-case, print only a
single page at a time. Subdividing the problem at the page level
would offer no performance improvement from the sequential
case.
A hybrid approach would be to subdivide the pages and process regions
of a page in a thread, as illustrated in Figure 3.6.
 Chapter 3: Fundamental Concepts of Parallel Programming 51

Multiple threads processing multiple page sections

Figure 3.6 Parallel Error Diffusion for Multi-thread, Multi-page Situation

Note that each thread must work on sections from different page.
This increases the startup latency involved before the threads can begin
work. In Figure 3.6, Thread 2 incurs a 1/3 page startup latency before it
can begin to process data, while Thread 3 incurs a 2/3 page startup
latency. While somewhat improved, the hybrid approach suffers from
similar limitations as the page-based partitioning scheme described
above. To avoid these limitations, you should focus on the row-based
error diffusion implementation illustrated in Figure 3.5.

Key Points
This chapter explored different types of computer architectures and how
they enable parallel software development. The key points to keep in
mind when developing solutions for parallel computing architectures are:
Decompositions fall into one of three categories: task, data, and
data flow.
Task-level parallelism partitions the work between threads based
on tasks.
Data decomposition breaks down tasks based on the data that the
threads work on.
52 Multi-Core Programming

Data flow decomposition breaks down the problem in terms of

how data flows between the tasks.
Most parallel programming problems fall into one of several well
known patterns.
The constraints of synchronization, communication, load balancing,
and scalability must be dealt with to get the most benefit out of a
parallel program.
Many problems that appear to be serial may, through a simple
transformation, be adapted to a parallel implementation.
 Chapter 4
Threading and
Parallel
Programming
Constructs
T his chapter describes the theory and practice of the principal parallel
programming constructs that focus on threading and begins with the
fundamental concepts of synchronization, critical section, and deadlock.
The following chapters cover implementation details and related issues.

Synchronization
Synchronization is an enforcing mechanism used to impose constraints
on the order of execution of threads. The synchronization controls the
relative order of thread execution and resolves any conflict among
threads that might produce unwanted behavior. In simple terms,
synchronization is used to coordinate thread execution and manage
shared data.
In an environment where messages are used for communicating
between a sender and a receiver, synchronization is implicit, as a
message must be sent before the message can be received. On the other
hand, for a shared-memory based environment, threads have no implicit
interdependency unless some constraints are imposed.
Two types of synchronization operations are widely used: mutual
exclusion and condition synchronization. In the case of mutual

53
54 Multi-Core Programming

exclusion, one thread blocks a critical sectiona section of code that

contains shared dataand one or more threads wait to get their turn to
enter into the section. This helps when two or more threads share the
same memory space and run simultaneously. The mutual exclusion is
controlled by a scheduler and depends on the granularity of the
scheduler. Condition synchronization, on the other hand, blocks a
thread until the system state specifies some specific conditions. The
condition synchronization allows a thread to wait until a specific
condition is reached. Figure 4.1 shows the generic representation of
synchronization.

Figure 4.1 Generic Representation of Synchronization Block inside Source

Code

While a number of techniques are available for synchronization, only

a few methods are used by developers on a regular basis. The techniques
used are also to some extent determined by the programming
environment.
 Chapter 4: Threading and Parallel Programming Constructs 55

The scope of synchronization is broad. Proper synchronization orders

the updates to data and provides an expected outcome. In Figure 4.2,
shared data d can get access by threads Ti and Tj at time ti, tj, tk, tl, where
ti tj tk tl and a proper synchronization maintains the order to update
d at these instances and considers the state of d as a synchronization
function of time. This synchronization function, s, represents the
behavior of a synchronized construct with respect to the execution time
of a thread.

Ti Tj

ti
tj

d T = f (t)
tk Shared Data d = f (t) = s(...,ti ,tj ,tk ,tl ,...)
tl

Shared data d depends on synchronization functions of time

Figure 4.2 Shared Data Synchronization, Where Data d Is Protected by a

Synchronization Operation

Figure 4.3 represents how synchronization operations are performed

in an actual multi-threaded implementation in a generic form, and
demonstrates the flow of threads. When m>=1, the creation timing for
initial threads T1Tm might not be the same. After block Bi as well as Bj,
the number of threads could be different, which means m is not
necessarily equal to n and n is not necessarily equal to p. For all
operational environments, the values of m, n, and p are at least 1.
56 Multi-Core Programming

Implementation Source Code Operational Flow of Threads

T1
T2 Tm

Perform synchronization
Parallel Code Block operations using parallel
or a section needs constructs Bi
multithread synchronization

T1 T2 Tn

Perform synchronization
operations using parallel
Parallel Code Block constructs Bj

T1...p

Figure 4.3 Operational Flow of Threads for an Application

Critical Sections
A section of a code block called a critical section is where shared
dependency variables reside and those shared variables have dependency
among multiple threads. Different synchronization primitives are used to
keep critical sections safe. With the use of proper synchronization
techniques, only one thread is allowed access to a critical section at any
one instance. The major challenge of threaded programming is to
implement critical sections in such a way that multiple threads perform
mutually exclusive operations for critical sections and do not use critical
sections simultaneously.
 Chapter 4: Threading and Parallel Programming Constructs 57

Critical sections can also be referred to as synchronization blocks.

Depending upon the way critical sections are being used, the size of a
critical section is important. Minimize the size of critical sections when
practical. Larger critical sections-based code blocks should split into multiple
code blocks. This is especially important in code that is likely to experience
significant thread contention. Each critical section has an entry and an exit
point. A critical section can be represented as shown in Figure 4.4.

<Critical Section Entry,

to keep other threads in waiting status>
...
Critical Section
...
<Critical Section Exit,
allow other threads to enter critical section>

Figure 4.4 Implement Critical Section in Source Code

Deadlock
Deadlock occurs whenever a thread is blocked waiting on a resource of
another thread that will never become available. According to the
circumstances, different deadlocks can occur: self-deadlock, recursive
deadlock, and lock-ordering deadlock. In most instances, deadlock means
lock-ordering deadlock.
The self-deadlock is the instance or condition when a thread, Ti, wants
to acquire a lock that is already owned by thread Ti. In Figure 4.5 (a), at
time ta thread Ti owns lock li, where li is going to get released at tc.
However, there is a call at tb from Ti, which requires li. The release time of
li is td, where td can be either before or after tc. In this scenario, thread Ti is
in self-deadlock condition at tb. When the wakeup path of thread Ti, resides
in another thread, Tj, that condition is referred to as recursive deadlock, as
shown in Figure 4.5 (b). Figure 4.5 (c) illustrates a lock-ordering thread,
where thread Ti locks resource rj and waits for resource ri, which is being
locked by thread Tj. Also, thread Tj locks resource ri and waits for resource
rj, which is being locked by thread Ti. Here, both threads Ti and Tj are in
deadlock at td, and w is the wait-function for a lock.
58 Multi-Core Programming

Ti Tj
Ti

ta

tb Section
Required
li by thread
Ti

td

tc

(a) Self-deadlock (b) Recursive Deadlock

Ti Tj

ti
tj

wi = f (td )

rj wj = f (td ) ri

td ti

(c) Lock-ordering Deadlock

Figure 4.5 Deadlock Scenarios

 Chapter 4: Threading and Parallel Programming Constructs 59

To represent the transition model of deadlock of an environment,

consider representing atomic states by si and each thread of the system
by Ti. Each thread can transition from one state to another by requesting
a resource, acquiring a resource, or freeing the current resource. So,
the transition can be represented as shown in Figure 4.6, where,
ri requesting a resource, ai acquiring a resource, and fi freeing
current resource.

Ti : ri ai fi sd
si sj

Current State Deadlock State

Figure 4.6 Deadlock Scenario in a State Transition for a Thread

For any thread Ti, if the state transition of Ti becomes sd for all
possible scenarios and remains blocked at sd, thread Ti would not have
any way to transition from sd to any other state. That is why state sd is
called the deadlock state for thread Ti.
Avoiding deadlock is one of the challenges of multi-threaded
programming. There must not be any possibility of deadlock in an
application. A lock-holding prevention mechanism or the creation of lock
hierarchy can remove a deadlock scenario. One recommendation is to
use only the appropriate number of locks when implementing
synchronization. Chapter 7 has a more detailed description of deadlock
and how to avoid it.

Synchronization Primitives
Synchronization is typically performed by three types of primitives:
semaphores, locks, and condition variables. The use of these primitives
depends on the application requirements. These synchronization primitives
are implemented by atomic operations and use appropriate memory fences.
A memory fence, sometimes called a memory barrier, is a processor
60 Multi-Core Programming

dependent operation that guarantees that threads see other threads memory
operations by maintaining reasonable order. To hide the granularity of these
synchronization primitives, higher level synchronizations are used. That way
application developers have to concern themselves less about internal
details.

Semaphores
Semaphores, the first set of software-oriented primitives to accomplish
mutual exclusion of parallel process synchronization, were introduced by
the well known mathematician Edsger Dijkstra in his 1968 paper, The
Structure of the THE-Multiprogramming System (Dijkstra 1968). Dijkstra
illustrated that synchronization can be achieved by using only traditional
machine instructions or hierarchical structure. He proposed that a
semaphore can be represented by an integer, sem, and showed that a
semaphore can be bounded by two basic atomic operations, P (proberen,
which means test) and V (verhogen, which means increment). These
atomic operations are also referred as synchronizing primitives. Even
though the details of Dijkstras semaphore representation have evolved,
the fundamental principle remains same. Where, P represents the
potential delay or wait and V represents the barrier removal or
release of a thread. These two synchronizing primitives can be
represented for a semaphore s as follows:
Thread "T" performs operation "P":
P(s) atomic {sem = sem-1; temp = sem}
if (temp < 0)
{Thread T blocked and enlists on a
waiting list for s}
Thread "T" performs operation "V":
V(s) atomic {sem = sem +1; temp = sem}
if (temp <=0)
{Release one thread from the waiting
list for s}
where semaphore value sem is initialized with the value 0 or 1 before the
parallel processes get started. In Dijkstras representation, T referred to
processes. Threads are used here instead to be more precise and to
remain consistent about the differences between threads and processes.
The P operation blocks the calling thread if the value remains 0, whereas
the V operation, independent of P operation, signals a blocked thread to
allow it to resume operation. These P and V operations are indivisible
actions and perform simultaneously. The positive value of sem
 Chapter 4: Threading and Parallel Programming Constructs 61

represents the number of threads that can proceed without blocking, and
the negative number refers to the number of blocked threads. When the
sem value becomes zero, no thread is waiting, and if a thread needs to
decrement, the thread gets blocked and keeps itself in a waiting list.
When the value of sem gets restricted to only 0 and 1, the semaphore is a
binary semaphore.
To use semaphore, you can consider semaphore as a counter, which
supports two atomic operations. Implementation of semaphores varies.
From a usability perspective, two kinds of semaphores exist: strong and
weak. These represent the success of individual calls on P. A strong
semaphore maintains First-Come-First-Serve (FCFS) model and provides
guarantee to threads to calls on P and avoid starvation. And a weak
semaphore is the one which does not provide any guarantee of service to
a particular thread and the thread might starve. For example, in POSIX,
the semaphores can get into starvation status and implemented
differently than what Dijkstra defined and considered as a weak
semaphore (Reek 2002).
According to Dijkstra, the mutual exclusion of parallel threads using
P and V atomic operations represented as follows:
semaphore s
s.sem = 1
begin
T: <non-critical section>
P(s)
<critical section>
V(s)
Goto T
end

Semaphores are largely of historical interest. They are the

unstructured goto statements of multi-threaded programming. Most
programming environments provide higher-level structured synchroniza-
tion primitives. However, like the goto, semaphores are occasionally the
best primitive to use. A typical use of a semaphore is protecting a shared
resource of which at most n instances are allowed to exist
simultaneously. The semaphore starts out with value n. A thread that
needs to acquire an instance of the resource performs operation P. It
releases the resource using operation V.
62 Multi-Core Programming

Lets examine how semaphores might be used for the producer-

consumer problem and whether the problem can be resolved using
semaphores or not. Producer-consumer is a classic synchronization
problem, also known as the bounded-buffer problem. Here a producer
function generates data to be placed in a shared buffer and a consumer
function receives the data out of the buffer and operates on it, where
both producer and consumer functions execute concurrently.
Pseudo-code using a semaphore for the producer-consumer problem
is shown in Figure 4.7.

semaphore s

void producer () {
while (1) {
<produce the next data>
s->release()
}
}

void consumer() {
while (1) {
s->wait()
<consume the next data>
}
}

Figure 4.7 Pseudo-code of Producer-Consumer Problem

Here neither producer nor consumer maintains any order. If the

producer function operates forever prior to the consumer function then
the system would require an infinite capacity and that is not possible. That
is why the buffer size needs to be within a boundary to handle this type of
scenario and make sure that if the producer gets ahead of the consumer
then the time allocated for the producer must be restricted. The problem
of synchronization can be removed by adding one more semaphores in the
previous solution shown in Figure 4.7. Adding the semaphore would
maintain the boundary of buffer as shown in Figure 4.8, where sEmpty and
sFull retain the constraints of buffer capacity for operating threads.
 Chapter 4: Threading and Parallel Programming Constructs 63

semaphore sEmpty, sFull

void producer() {
while (1) {
sEmpty->wait()
<produce the next data>
sFull->release()
}
}

void consumer() {
while (1) {
sFull->release()
<consume the next data>
sEmpty->wait()
}
}

Figure 4.8 Dual Semaphores Solution for Producer-Consumer Problem

Instead of using two independent semaphores and having a

constraint-based solution, the solution in Figure 4.8 can be implemented
using other synchronization primitives as well. The following sections
discuss how to solve the producer-consumer problem using locks and
conditional variables primitives.

Locks
Locks are similar to semaphores except that a single thread handles a lock
at one instance. Two basic atomic operations get performed on a lock:
acquire(): Atomically waits for the lock state to be unlocked and
sets the lock state to lock.
release(): Atomically changes the lock state from locked to
unlocked.
At most one thread acquires the lock. A thread has to acquire a lock
before using a shared resource; otherwise it waits until the lock becomes
available. When one thread wants to access shared data, it first acquires
the lock, exclusively performs operations on the shared data and later
releases the lock for other threads to use. The level of granularity can be
either coarse or fine depending on the type of shared data that needs to
be protected from threads. The coarse granular locks have higher lock
contention than finer granular ones. To remove issues with lock
64 Multi-Core Programming

granularity, most of the processors support the Compare and Swap (CAS)
operation, which provides a way to implement lock-free synchronization.
The atomic CAS operations guarantee that the shared data remains
synchronized among threads. If you require the use of locks, it is
recommended that you use the lock inside a critical section with a single
entry and single exit, as shown in Figure 4.9.

{define all necessary locks}

<Start multithreading blocks>
...
<critical section start>

<acquire lock L>

.. operate on shared memory protected by lock L ..

<release lock L>

<critical section end>
...
<End multithreading blocks>

Figure 4.9 A Lock Used Inside a Critical Section

From an implementation perspective, it is always safe to use explicit

locks rather than relying on implicit locks. In general a lock must not be
held for a long periods of time. The explicit locks are defined by the
developer, whereas implicit locks come from the underlying framework
used, such as database engines provides lock the maintain data
consistency.
In the produce-consumer problem, if the consumer wants to
consume a shared data before the producer produces, it must wait. To
use locks for the producer-consumer problem, the consumer must loop
until the data is ready from the producer. The reason for looping is that
the lock does not support any wait operation, whereas Condition
Variables does.

Lock Types
An application can have different types of locks according to the
constructs required to accomplish the task. You must avoid mixing lock
types within a given task. For this reason, special attention is required
 Chapter 4: Threading and Parallel Programming Constructs 65

when using any third party library. If your application has some third
party dependency for a resource R and the third party uses lock type L
for R, then if you need to use a lock mechanism for R, you must use lock
type L rather any other lock type. The following sections cover these
locks and define their purposes.
Mutexes. The mutex is the simplest lock an implementation can use.
Some texts use the mutex as the basis to describe locks in general. The
release of a mutex does not depend on the release() operation only. A
timer attribute can be added with a mutex. If the timer expires before a
release operation, the mutex releases the code block or shared memory
to other threads. A try-finally clause can be used to make sure that the
mutex gets released when an exception occurs. The use of a timer or try-
finally clause helps to prevent a deadlock scenario.
Recursive Locks. Recursive locks are locks that may be repeatedly
acquired by the thread that currently owns the lock without causing the
thread to deadlock. No other thread may acquire a recursive lock until
the owner releases it once for each time the owner acquired it. Thus
when using a recursive lock, be sure to balance acquire operations with
release operations. The best way to do this is to lexically balance the
operations around single-entry single-exit blocks, as was shown for
ordinary locks. The recursive lock is most useful inside a recursive
function. In general, the recursive locks are slower than nonrecursive
locks. An example of recursive locks use is shown in Figure 4.10.

Recursive_Lock L
void recursiveFunction (int count) {
L->acquire()
if (count > 0) {
count = count - 1;
recursiveFunction(count);
}
L->release();
}

Figure 4.10 An Example of Recursive Lock Use

Read-Write Locks. Read-Write locks are also called shared-exclusive or

multiple-read/single-write locks or non-mutual exclusion semaphores.
Read-write locks allow simultaneous read access to multiple threads but
limit the write access to only one thread. This type of lock can be used
66 Multi-Core Programming

efficiently for those instances where multiple threads need to read shared
data simultaneously but do not necessarily need to perform a write
operation. For lengthy shared data, it is sometimes better to break the
data into smaller segments and operate multiple read-write locks on
the dataset rather than having a data lock for a longer period of time.
Spin Locks. Spin locks are non-blocking locks owned by a thread.
Waiting threads must spin, that is, poll the state of a lock rather than
get blocked. Spin locks are used mostly on multiprocessor systems. This
is because while the thread spins in a single-core processor system, no
process resources are available to run the other thread that will release
the lock. The appropriate condition for using spin locks is whenever the
hold time of a lock is less than the time of blocking and waking up a
thread. The change of control for threads involves context switching of
threads and updating thread data structures, which could require more
instruction cycles than spin locks. The spin time of spin locks should be
limited to about 50 to 100 percent of a thread context switch (Kleiman
1996) and should not be held during calls to other subsystems. Improper
use of spin locks might cause thread starvation. Think carefully before
using this locking mechanism. The thread starvation problem of spin
locks can be alleviated by using a queuing technique, where every
waiting thread to spin on a separate local flag in memory using First-In,
First-Out (FIFO) or queue construct.

Condition Variables
Condition variables are also based on Dijkstras semaphore semantics,
with the exception that no stored value is associated with the operation.
This means condition variables do not contain the actual condition to
test; a shared data state is used instead to maintain the condition for
threads. A thread waits or wakes up other cooperative threads until a
condition is satisfied. The condition variables are preferable to locks
when pooling requires and needs some scheduling behavior among
threads. To operate on shared data, condition variable C, uses a lock, L.
Three basic atomic operations are performed on a condition variable C:
wait(L): Atomically releases the lock and waits, where wait
returns the lock been acquired again
signal(L): Enables one of the waiting threads to run, where signal
returns the lock is still acquired
broadcast(L): Enables all of the waiting threads to run, where
broadcast returns the lock is still acquired
 Chapter 4: Threading and Parallel Programming Constructs 67

To control a pool of threads, use of a signal function is recommended.

The penalty for using a broadcast-based signaling function could be
severe and extra caution needs to be undertaken before waking up all
waiting threads. For some instances, however, broadcast signaling can be
effective. As an example, a write lock might allow all readers to
proceed at the same time by using a broadcast mechanism.
To show the use of a condition variable for a synchronization
problem, the pseudocode in Figure 4.11 solves the producer-consumer
problem discussed earlier. A variable LC is used to maintain the
association between condition variable C and an associated lock L.

Condition C;
Lock L;
Bool LC = false;

void producer() {
while (1) {
L ->acquire();
// start critical section
while (LC == true) {
C ->wait(L);
}
// produce the next data
LC = true;
C ->signal(L);
// end critical section
L ->release();
}
}

void consumer() {
while (1) {
L ->acquire();
// start critical section
while (LC == false) {
C ->wait(L);
}
// consume the next data
LC = false;
// end critical section
L ->release();
}
}

Figure 4.11 Use of a Condition Variable for the Producer-Consumer Problem

68 Multi-Core Programming

Monitors
For structured synchronization, a higher level construct is introduced
for simplifying the use of condition variables and locks, known as a
monitor. The purpose of the monitor is to simplify the complexity of
primitive synchronization operations and remove the implementation
details from application developers. The compiler for the language that
supports monitors automatically inserts lock operations at the
beginning and the end of each synchronization-aware routine. Most
recent programming languages do not support monitor explicitly,
rather they expose lock and unlock operations to the developers. The
Java language supports explicit monitor objects along with
synchronized blocks inside a method. In Java, the monitor is maintained
by the synchronized constructs, such as
synchronized (object) {

<Critical Section>
}

where the condition primitives are used by wait(), notify(), or

notifyAll() methods. Do not confuse this with the Monitor object in
the Java SDK though. The Java Monitor object is used to perform
resource management in Java Management Extension (JMX). Similarly,
the monitor object in C# is used as lock construct.

Messages
The message is a special method of communication to transfer
information or a signal from one domain to another. The definition of
domain is different for different scenarios. For multi-threading
environments, the domain is referred to as the boundary of a thread.
The three Ms of message passing are multi-granularity,
multithreading, and multitasking (Ang 1996). In general, the
conceptual representations of messages get associated with processes
rather than threads. From a message-sharing perspective, messages get
shared using an intra-process, inter-process, or process-process
approach, as shown in Figure 4.12.
 Chapter 4: Threading and Parallel Programming Constructs 69

Figure 4.12 Message Passing Model

Two threads that communicate with messages and reside in the same
process use intra-process messaging. Two threads that communicate and
reside in different processes use inter-process messaging. From the
developers perspective, the most common form of messaging is the
process-process approach, when two processes communicate with each
other rather than depending on the thread.
In general, the messaging could be devised according to the memory
model of the environment where the messaging operation takes place.
Messaging for the shared memory model must be synchronous, whereas
for the distributed model messaging can be asynchronous. These
operations can be viewed at a somewhat different angle. When there is
nothing to do after sending the message and the sender has to wait for
the reply to come, the operations need to be synchronous, whereas if the
sender does not need to wait for the reply to arrive and in order to
proceed then the operation can be asynchronous.
The generic form of message communication can be represented as
follows:
Sender:
<sender sends message to one or more recipients
through structure>
\\ Here, structure can be either queue or port
<if shared environment>
{wait for the acknowledgement>
<else>
{sender does the next possible operation>
70 Multi-Core Programming

Receiver:
<might wait to get message from sender from
appropriate structure>
<receive message from appropriate structure and
process>
The generic form of message passing gives the impression to developers
that there must be some interface used to perform message passing. The
most common interface is the Message Passing Interface (MPI). MPI is
used as the medium of communication, as illustrated in Figure 4.13.

MPI Communication method

MPI Interface

MPI Interface
over base network protocol

Ni Nj

N and Nj are two diferent nodes/processes

i

Figure 4.13 Basic MPI Communication Environment

To synchronize operations of threads, semaphores, locks, and

condition variables are used. These synchronization primitives convey
status and access information. To communicate data, they use thread
messaging. In thread messaging, synchronization remains explicit, as
there is acknowledgement after receiving messages. The
acknowledgement avoids primitive synchronization errors, such as
deadlocks or race conditions. The basic operational concepts of
messaging remain the same for all operational models. From an
implementation point of view, the generic client-server model can be
used for all messaging models.
Inside hardware, message processing occurs in relationship with the
size of the message. Small messages are transferred between processor
registers and if a message is too large for processor registers, caches get
used. Larger messages require main memory. In the case of the largest
messages, the system might use processor-external DMA, as shown in
Figure 4.14.
 Chapter 4: Threading and Parallel Programming Constructs 71

Microprocessor

Register
Use of hardware
system components
according to the size Caches (L0, L1 <L2=LLC>)
of messages

DMA Hardware / Processor

Also known as I/O Processor, external to main processor

Figure 4.14 System Components Associated with Size of Messages

Flow Control-based Concepts

In the parallel computing domain, some restraining mechanisms allow
synchronization among multiple attributes or actions in a system. These
are mainly applicable for shared-memory multiprocessor or multi-core
environments. The following section covers only two of these concepts,
fence and barrier.

Fence
The fence mechanism is implemented using instructions and in fact, most
of the languages and systems refer to this mechanism as a fence
instruction. On a shared memory multiprocessor or multi-core
environment, a fence instruction ensures consistent memory operations.
At execution time, the fence instruction guarantees completeness of all
pre-fence memory operations and halts all post-fence memory operations
until the completion of fence instruction cycles. This fence mechanism
ensures proper memory mapping from software to hardware memory
models, as shown in Figure 4.15. The semantics of the fence instruction
depend on the architecture. The software memory model implicitly
supports fence instructions. Using fence instructions explicitly could be
error-prone and it is better to rely on compiler technologies. Due to the
performance penalty of fence instructions, the number of memory fences
needs to be optimized.
72 Multi-Core Programming

Figure 4.15 Fence Mechanism

Barrier
The barrier mechanism is a synchronization method by which threads
in the same set keep collaborating with respect to a logical
computational point in the control flow of operations. Through this
method, a thread from an operational set has to wait for all other
threads in that set to complete in order to be able to proceed to the
next execution step. This method guarantees that no threads proceed
beyond an execution logical point until all threads have arrived at that
logical point. Barrier synchronization is one of the common operations
for shared memory multiprocessor and multi-core environments. Due to
the aspect of waiting for a barrier control point in the execution flow,
the barrier synchronization wait function for ith thread can be
represented as
(Wbarrier)i = f ((Tbarrier)i, (Rthread)i)

where Wbarrier is the wait time for a thread, Tbarrier is the number of threads
has arrived, and Rthread is the arrival rate of threads.
For performance consideration and to keep the wait time within a
reasonable timing window before hitting a performance penalty, special
 Chapter 4: Threading and Parallel Programming Constructs 73

consideration must be given to the granularity of tasks. Otherwise, the

implementation might suffer significantly.

Implementation-dependent Threading Features

The functionalities and features of threads in different environments are
very similar; however the semantics could be different. That is why
the conceptual representations of threads in Windows and Linux remain
the same, even though the way some concepts are implemented could
be different. Also, with the threading APIs of Win32, Win64, and POSIX
threads (Pthreads), the semantics are different as well. Windows
threading APIs are implemented and maintained by Microsoft and
work on Windows only, whereas the implementation of Pthreads APIs
allows developers to implement threads on multiple platforms. The IEEE
only defined the Pthreads APIs and let the implementation be done by OS
developers. Due to the implementation issues of Pthreads, not all features
exist in Pthreads APIs. Developers use Pthreads as a wrapper of their
own thread implementations. There exists a native Linux Pthreads library
similar to Windows native threads, known as Native POSIX Thread
Library (NPTL).
Consider the different mechanisms used to signal threads in Windows
and in POSIX threads. Windows uses an event model to signal one or
more threads that an event has occurred. However, no counterpart to
Windows events is implemented in POSIX threads. Instead, condition
variables are used for this purpose.
These differences are not necessarily limited to cross-library
boundaries. There may be variations within a single library as well. For
example, in the Windows Win32 API, Microsoft has implemented two
versions of a mutex. The first version, simply referred to as a mutex,
provides one method for providing synchronized access to a critical
section of the code. The other mechanism, referred to as a
CriticalSection, essentially does the same thing, with a completely
different API. Whats the difference?
The conventional mutex object in Windows is a kernel mechanism.
As a result, it requires a user-mode to kernel-mode transition to work.
This is an expensive operation, but provides a more powerful
synchronization mechanism that can be used across process boundaries.
However, in many applications, synchronization is only needed within a
single process. Therefore, the ability for a mutex to work across process
boundaries is unnecessary, and leads to wasted overhead. To remove
74 Multi-Core Programming

overhead associated with the standard mutex Microsoft implemented the

CriticalSection, which provides a user-level locking mechanism. This
eliminates any need to make a system call, resulting in much faster
locking.
Even though different threading libraries will have different ways of
implementing the features described in this chapter, the key to being
able to successfully develop multithreaded applications is to understand
the fundamental concepts behind these features.

Key Points
This chapter presented parallel programming constructs and later
chapters provide details about the implementation of the constructs. To
become proficient in threading techniques and face fewer issues during
design and development of a threaded solution, an understanding of the
theory behind different threading techniques is helpful. Here are some of
the points you should remember:
For synchronization, an understanding of the atomic actions of
operations will help avoid deadlock and eliminate race
conditions.
Use a proper synchronization construct-based framework for
threaded applications.
Use higher-level synchronization constructs over primitive types.
An application cannot contain any possibility of a deadlock
scenario.
Threads can perform message passing using three different
approaches: intra-process, inter-process, and process-process.
Understand the way threading features of third-party libraries are
implemented. Different implementations may cause applications
to fail in unexpected ways.
 Chapter 5
Threading APIs
revious chapters introduced the basic principles behind writing
P concurrent applications. While describing every thread package
available today is beyond the scope of this book, it is important to
illustrate the aforementioned principles with practical examples. This
chapter will provide an overview of several popular thread packages
used by developers today.

Threading APIs for Microsoft Windows

Since the advent of Windows NT, Microsoft has enabled application
developers to write multi-threaded software applications. Advances in
processor technology have allowed Microsoft to add additional
capabilities, evolving its programming interface to support more and
more advanced threading operations. As a result, the Microsoft threading
APIs have become a powerful tool for writing multi-threaded applications
running on Windows.
Historically, native Windows applications have been written in
C/C++ using Microsofts Win32 or MFC APIs. However, many new
applications are being developed using Microsofts .NET platform and
associated common language runtime (CLR). This chapter examines
writing multi-threaded applications using both programming
techniques.

Win32/MFC Thread APIs

The Win32/MFC API provides developers with a C/C++ interface for
developing Windows applications. While there are a large number of

75
76 Multi-Core Programming

developers moving to a managed code development environment,

understanding the Win32/MFC model is important for several reasons:
Performance. Applications that run in a managed runtime
environment run on a software virtual machine. Managed code is
compiled to virtual machine op-codes. These op-codes are then
translated, at runtime, to native processor instructions. A number of
optimizations minimize this overhead. However, native applications
run directly on the processor; the overhead of translating virtual
machine op-codes is eliminated. In some applications, the overhead
of the virtual machine may not be acceptable.
Legacy Application Support. There are a number of existing
applications where it is not reasonable to port to a managed
runtime environment. In order to maximize the performance of
these applications for multi-core processors, the existing code
must be multi-threaded.

Creating Threads
All processes start with a single thread of execution: the main thread. In
order to write multi-threaded code, one must have the ability to create
new threads. The most basic thread creation mechanism provided by
Microsoft is CreateThread():
HANDLE CreateThread(
LPSECURITY_ATTRIBUTES lpThreadAttributes,
SIZE_T dwStackSize,
LPTHREAD_START_ROUTINE lpStartAddress,
LPVOID lpParameter,
DWORD dwCreationFlags,
LPDWORD lpThreadId );
The first parameter, lpThreadAttributes, is a data structure that
specifies several different security parameters. It also defines whether or
not processes created from the current process (child processes) inherit
this handle. In other words, this parameter gives advanced control over
how the thread handle may be used in the system. If the programmer
does not need control over these attributes, the programmer may specify
NULL for this field.
The second parameter, dwStackSize, specifies the stack size of the
thread. The size is specified in bytes, and the value is rounded up to the
nearest page size.
 Chapter 5: Threading APIs 77

The third parameter, lpStartAddress, specifies a function pointer

to the actual code that the thread runs. The function pointer that has the
following signature:
DWORD WINAPI ThreadFunc(LPVOID data)
From the thread functions signature, one can see that a thread will
return a status value on exit and will take a void pointer to some data
value or structure. This provides the basic communication mechanism
between the thread and the outside world.
The fourth parameter in CreateThread(), lpParameter, is the
data value to pass into the thread function. In other words, considering
the aforementioned ThreadFunc(), the value specified in this
argument to CreateThread() will be passed in as the data value in
ThreadFunc().
The fifth parameter, dwCreationFlags, specifies various
configuration options. For example, using this flag, the programmer may
specify that the thread be created in the suspended state, thus giving the
programmer control of when the thread is started.
The final parameter, lpThreadId, points to where the function
should store a value that uniquely identifies a thread in the system. This
identifier is global and is useful in debugging.
On success, CreateThread() returns a HANDLE to the new thread.
An observant reader may at this point ask what the difference is between
the HANDLE returned by CreateThread() and the Thread ID. There are
actually a number of differences; for our purposes its only important to
note that there are two different return values for CreateThread() and
that different Thread API calls will expect one value or the other. The
HANDLE value is the most frequently used of the two.
Once a thread is created, the programmer will at some point want to
terminate the thread. This is accomplished by calling the ExitThread()
function:
VOID ExitThread(DWORD dwExitCode);
ExitThread() is called at the end of ThreadFunc() to indicate that
the thread should terminate. Note that an explicit call to ExitThread()
is not mandatory; simply returning the exit code will result in this
function being called implicitly:
DWORD WINAPI ThreadFunc( LPVOID data )
{
78 Multi-Core Programming

// do something
...
// ready to exit thread
return 0; // will implicitly call ExitThread(0);
}
Note that in C++, calling ExitThread() will exit the thread before
any constructors/automatic variables are cleaned up. Thus, Microsoft
recommends that the program simply return from the ThreadFunc()
rather than call ExitThread() explicitly.
The CreateThread() and ExitThread() functions provide a
flexible, easy to use mechanism for creating threads in Windows
applications. Theres just one problem. CreateThread() does not
perform per-thread initialization of C runtime datablocks and variables.
Hence, you cannot use CreateThread() and ExitThread(), in any
application that uses the C runtime library. Instead, Microsoft provides
two other methods, _beginthreadex() and _endthreadex() that
perform the necessary initialization prior to calling CreateThread().
CreateThread() and ExitThread() are adequate for writing
applications that just use the Win32 API; however, for most cases, it is
recommended that developers use _beginthreadex() and
_endthreadex() to create threads.
The definition of _beginthreadex() is similar to that of
CreateThread(); the only difference being one of semantics.
unsigned long _beginthreadex( // unsigned long
// instead of HANDLE,
// but technically the
// same
void *security, // same as CreateThread()
unsigned stack_size, // same as CreateThread()
unsigned (__stdcall func) (void), // ptr to func
// returning unsigned
// instead of void
void *arglist, // same as CreateThread()
unsigned initflag, // same as CreateThread()
unsigned* threadID); // same as CreateThread()
Similarly, the definition of _endthreadex() follows that of
ExitThread():
void _endthreadex( unsigned retval );
 Chapter 5: Threading APIs 79

For applications written in C++ using MFC, Microsoft provides yet

1
another function to create a threadAfxBeginThread():
CWinThread* AfxBeginThread(
AFX_THREADPROC pfnThreadProc,
LPVOID pParam,
int nPriority = THREAD_PRIORITY_NORMAL,
UINT nStackSize = 0,
DWORD dwCreateFlags = 0,
LPSECURITY_ATTRIBUTES lpSecurityAttrs = NULL );
AfxBeginThread() differs from CreateThread() in several ways:
AfxBeginThread() returns a pointer to a CWinThread object
rather than a HANDLE to a thread.
AfxBeginThread() re-orders the parameters and replaces the
threadId parameter with nPriority, which allows the
programmer to specify the threads priority.
AfxBeginThread() expects the following definition for
ThreadFunc():
UINT ThreadFunc (LPVOID pParam);
AfxBeginThread() calls _beginthreadex(); thus it is safe to
use with the C runtime library.
For all intent and purpose, AfxBeginThread is conceptually identical to
CreateThread. MFC provides a complimentary function to ExitThread
as well:
void AFXAPI AfxEndThread( UINT nExitCode,
BOOL bDelete = TRUE );
The bDelete parameter specifies whether or not the framework
should automatically delete the associated thread object upon
termination. It should be set to FALSE if a program wishes to check the
exit code of the thread; however, it then becomes the programs
responsibility to destroy the CWinThread object.

Managing Threads
Now that you know how to create a thread, lets examine the process of
controlling or manipulating the execution of threads. It was previously
demonstrated that Windows allows developers to create threads in one
of two initial states: suspended or running. For the remainder of the

1
There are two types of threads that can be created using AfxBeginThread(): worker threads and
user-interface threads. This text only considers worker threads.
80 Multi-Core Programming

chapter, well use Win32 definitions to illustrate the concepts of

programming in Windows. Most MFC calls will be identical to Win32; the
only difference is that the MFC calls will be methods called on MFC-based
classes, such as CWinThread, rather than C function calls.
The following functions allow the programmer to control the
execution of a thread:
DWORD SuspendThread( HANDLE hThread );
DWORD ResumeThread( HANDLE hThread );
BOOL TerminateThread( HANDLE hThread, DWORD dwExitCode );

SuspendThread() allows the developer to suspend execution of the

thread specified by the HANDLE parameter. The kernel keeps track of
the current suspend count for a given thread in the threads data
structure. A suspend count of 0 indicates that the thread is ready to run.
A suspend count greater than 0 indicates that the thread is suspended.
SuspendThread(), when called, will increment this field and return the
previous value of suspend count. ResumeThread() will decrement the
suspend count for the thread specified by the HANDLE value. It will
return the previous suspend count. This implies that if a thread is
transitioning from the suspended state to the run state, its suspend count
will be 1. Calling ResumeThread() on a currently running thread will
return 0. This does not indicate an error condition, as calling
ResumeThread() on a thread with a suspend count of 0 has no effect.
The TerminateThread() function forces the thread specified by the
HANDLE parameter to terminate. No user code executes; the thread is
immediately terminated. If the function is successful, a non-zero value is
returned.
Developers must be very careful when calling SuspendThread(),
as the thread may be in a state in which it is dangerous to suspend.
For example, if the thread is holding a semaphore and is suspended, it
will not release the semaphore prior to being suspended. As a result,
other threads will not be able to access critical sections until the
suspended thread is resumed and releases the resource. This may
cause significant performance problems or even deadlock.
TerminateThread() is even more dangerous. The thread that is
being terminated will not be given a chance to do any clean-up;
therefore, a number of particularly nasty side-effects may occur. For
example, if a thread is holding on to a synchronization object such as a
mutex, and is abruptly terminated by this call, the synchronization object
will remain locked; hence, a deadlock will occur as the thread no longer
 Chapter 5: Threading APIs 81

exists and will be unable to release this resource. It is strongly

recommended that you avoid using this function.
In order to safely suspend/terminate threads, we need a signaling
mechanism that allows one thread, say the main thread, to notify the
targeted thread that it must suspend/terminate itself. Fortunately,
Windows provides developers with a way to do such an operation by
using Windows events.

Thread Communication using Windows Events

As previously demonstrated, multiple threads within an application need
a mechanism that can be used for inter-thread communication. Microsoft
has provided Event objects that may be used for this purpose. The code
sample provided in Listing 5.1 illustrates the use of Windows Events to
communicate between threads.

1 // This example illustrates the use of

2 // Windows Events as a inter-thread communication
3 // mechanism.
4 #define NUM_THREADS 10
5 #include <windows.h>
6 #include <stdio.h>
7 #include <process.h>
8
9 typedef struct
10 {
11 int Id;
12 HANDLE hTerminate;
13 } ThreadArgs;
14
15 unsigned __stdcall ThreadFunc( void *pArgs )
16 {
17 HANDLE hTerminate = ((ThreadArgs *)pArgs)->hTerminate;
18 int id = ((ThreadArgs *)pArgs)->Id;
19
20 // run until we are told to terminate
21 while (1)
22 {
23 // Check to see if we should terminate
24 if (WaitForSingleObject(hTerminate, 0) ==
25 WAIT_OBJECT_0)
26 {
27 // Terminate Thread - we call ResetEvent to
28 // return the terminate thread to its non-
29 // signaled state, then exit the while() loop
30 printf("Terminating Thread %d\n", id);
82 Multi-Core Programming

31 ResetEvent(hTerminate);
32 break;
33 }
34 // we can do our work now...
35 // simulate the case that it takes 1 s
36 // to do the work the thread has to do
37 Sleep(1000);
38 }
39 _endthreadex(0);
40 return 0;
41 }
42
43
44 int main( int argc, char* argv[] )
45 {
46 unsigned int threadID[NUM_THREADS];
47 HANDLE hThread[NUM_THREADS];
48 ThreadArgs threadArgs[NUM_THREADS];
49
50 // Create 10 threads
51 for (int i = 0; i < NUM_THREADS; i++)
52 {
53 threadArgs[i].Id = i;
54 threadArgs[i].hTerminate = CreateEvent(NULL, TRUE,
55 FALSE, NULL);
56 hThread[i] = (HANDLE)_beginthreadex(NULL, 0,
57 &ThreadFunc, &threadArgs[i], 0, &threadID[i]);
58 }
59
60 printf("To kill a thread (gracefully), press 0-9, " \
61 "then <Enter>.\n");
62 printf("Press any other key to exit.\n");
63
64 while (1)
65 {
66 int c = getc(stdin);
67 if (c == '\n') continue;
68 if (c < '0' || c > '9') break;
69 SetEvent(threadArgs[c '0'].hTerminate);
70 }
71
72 return 0;
73 }

Listing 5.1 A Thread Application that uses Windows Events

The application in Listing 5.1 is very simple. When loaded, the

application creates multiple threads to process different tasks. It uses
 Chapter 5: Threading APIs 83

_beginthreadex() as discussed in the previous section. The application

notifies the user that the application is started and gives the user an
interface in which a thread may be terminated. If the user enters a thread
ID, the application, using Windows Events, terminates the thread
gracefully. Otherwise, the different threads in the system continue to run
until the user indicates that the program should be terminated.
In order to use events, a programmer must first create an event. We
do this in line 54 using the CreateEvent() method:
HANDLE CreateEvent(
LPSECURITY_ATTRIBUTES lpEventAttributes,
BOOL bManualReset,
BOOL bInitialState,
LPCTSTR lpName );
The first parameter, lpEventAttributes, should look familiar. Its
the same security attributes data structure defined in the
CreateThread() function call. The default attributes are valid for this
case, so we simply pass in NULL. The second parameter,
bManualReset, allows the programmer to specify whether or not the
event being created should be explicitly reset by the programmer using
the ResetEvent function call. This parameter gives the programmer
the option to determine whether or not more than one thread will
respond to a given event. If bManualReset is FALSE, then Windows
will create an auto-reset event and return the event to the non-signaled
state after a single thread has been run. If bManualReset is TRUE,
Windows will create a manual reset event and it is up to the program to
return the event object to the non-signaled state. In the example given
in Listing 5.1, it was a moot point, as each thread had its own event that
it was monitoring. Therefore, the event could have just as easily been
set to FALSE. The third parameter, bInitialState, determines the
initial state of the Event. The programmer may specify that the event is
in either the signaled or non-signaled state. In the example application,
the event was initialized to the non-signaled state, indicating that the
user is not yet ready to terminate the thread. Finally, the programmer
may specify a name for the event in the fourth parameter, lpName.
Providing a name creates a system-wide event. Upon successful
completion, a handle to the event is returned. This handle is used in the
other API calls to operate on the event.
Once the event is created, it is ready to use. The event handle is stored
locally from the main thread, and is also passed to each individual thread.
This defines an interface between the threads that contain a reference to
this object. In this example, the main thread and individual worker threads
84 Multi-Core Programming

now have a communication channel. How do the threads communicate?

They communicate by using the messages described in Chapter 4. In
this case, our messages are Windows Events. By using SetEvent(),
ResetEvent(), and the Wait routines, WaitForSingleObject() and
WaitForMultipleObjects(), we can send messages between threads.
The Wait routines wait on a specific handleor multiple handles in the
case of WaitForMultipleObjects()for a specified timeout period. Since
a handle in Windows is a generic concept used to reference objects of
multiple types, the Wait routines provided by Microsoft wait on multiple
2
object types. In this aforementioned example, the handle used references
that an event object created specifically to terminate the thread. The timeout
period is 0, which indicates that the programmer is only interested in
checking to see if the event has been signaled, or a numeric value that
specifies a timeout. If the programmer specifies a timeout of 0, and the event
hasnt been signaled, WaitForSingleObject() will return immediately and
indicate to the programmer that the event has not yet occurred. In other
situations, it may make sense to specify a non-zero timeout value. In that case,
WaitForSingleObject() will wait for the period of time specified by the
timeout value for the event to occur. Microsoft defines a special constant,
INFINITE, to indicate that the thread of control wants to wait indefinitely for
the event to occur. In this way, the programmer can notify the OS that the
thread has no other work to do until this particular event occurs, and hence
can be moved off the run queue to the wait queue. The OS can then switch to
a thread that is in the ready-to-run state.
The function prototype for the WaitForSingleObject() function
has the following syntax:
DWORD WaitForSingleObject( HANDLE hHandle,
DWORD dwMilliseconds );
WaitForSingleObject() will return one of four values:
WAIT_OBJECT_0. This value is returned when the object that is
3
being waited on enters the signaled state.
WAIT_TIMEOUT. This value is returned when the specified
timeout value occurs prior to the object entering the signaled
state.

2
WaitForXXX() may wait on events, jobs, mutexes, processes, semaphores, threads, and timers,
among other objects.
3
The meaning of a signaled state varies based on the type of object being waited on. In the example
in Figure 5.1, we wait on an Event object, hence, WAIT_OBJECT_0 is returned once SetEvent()
sets the events state to signaled.
 Chapter 5: Threading APIs 85

WAIT_ABANDONED. In the case that the handle refers to a

Mutex object, this return code indicates that the thread that
owned the mutex did not release the mutex prior to
termination.
WAIT_FAILED. This value indicates that an error occurred.
GetLastError() should be used to get extended error
information regarding the cause of failure.
The function WaitForMultipleObjects() has the following
prototype:
DWORD WaitForMultipleObjects( DWORD nCount,
const HANDLE* lpHandles,
BOOL bWaitAll,
DWORD dwMilliseconds );
Note that the WaitForMultipleObjects() call takes in a different
parameter set than WaitForSingleObject(). The parameter nCount
specifies the number of handles to wait on. This value cannot exceed
the maximum number of object handles specified by the
MAXIMUM_WAIT_OBJECTS constant. The parameter lpHandles
specifies an array of object handles to wait on. Parameter bWaitAll
indicates whether or not the programmer wants to wait on all handles to
be signaled before returning, or wait on any one or more of the handles
to be signaled before returning. In the case of the former, the developer
should set bWaitAll to TRUE; for the latter case, the developer should
set bWaitAll to FALSE. The timeout value is the same for both
WaitForSingleObject() and WaitForMultipleObjects().
The return value for WaitForMultipleObjects() is identical in the
case of WAIT_TIMEOUT or WAIT_FAILED. In the case that an event is
signaled, or a handle is abandoned, the return value is slightly different.
In that case, WaitForMultipleObjects() returns WAIT_OBJECT_I or
WAIT_ABANDONED_I, where I is the index position in the array of
4
object handles where the signaled event was found. For example,
assuming that the programmer wanted to be notified when any object
was signaled, the code excerpt illustrated in Listing 5.2 can be used to
determine which object handle has been signaled.

4
If bWaitAll is set to FALSE, and if the number of objects in the signaled state happens to be greater
than 1, the array index of the first signaled or abandoned value in the arraystarting at array index
0is returned.
86 Multi-Core Programming

1 DWORD event, arrayIndex = 0;

2
3 // Assume eventArray and count are initialized elsewhere
4 // Wait for any of the events in eventArray to occur
5
6 event = WaitForMultipleObjects(
7 count,
8 eventArray,
9 FALSE,
10 INFINITE );
11 switch (event)
12 {
13 case WAIT_OBJECT_0 + 0:
14 // eventArray[0] signaled
15 case WAIT_OBJECT_0 + 1:
16 // eventArray[1] signaled
17 ...

Listing 5.2 Computing the Index of the Event that Has Been Signaled while Waiting
on Multiple Objects

Now that the thread has a mechanism for waiting for a particular
event to occur, we need a mechanism to signal the thread when it is time
to terminate. Microsoft provides the SetEvent() call for this purpose.
SetEvent() sets an event object to the signaled state. This allows a
thread to notify another thread that the event has occurred. SetEvent()
has the following signature:
BOOL SetEvent( HANDLE hEvent );
SetEvent() takes a single parameter which is the HANDLE value of
the specific event object, and returns TRUE if the event was signaled
successfully. The handle to the event object must be modifiable; in
other words, the access rights for the handle must have the
EVENT_MODIFY_STATE field set.
In the case of a manual reset event, the programmer must return the
event object the non-signaled state. To do this, a programmer uses the
ResetEvent() function. The ResetEvent() function has the following
prototype:
BOOL ResetEvent( HANDLE hEvent );
ResetEvent() accepts as a parameter the handle to reset and
returns TRUE upon success. Like SetEvent(), the handle to the event
 Chapter 5: Threading APIs 87

object must have the appropriate access rights set, otherwise the call to
5
ResetEvent() will fail.
It is important to contrast the example program in Listing 5.1 to the
case where the TerminateThread() function is used to terminate a
thread. TerminateThread() fails to give the thread any chance of
graceful exit; the thread is terminated immediately and without any
chance to properly free any resources it may have acquired. It
recommended that you use a notification mechanism such as the one
defined above to give the thread a chance to do proper cleanup.

Thread Synchronization
Generally speaking, creating a thread is a relatively simple task, and one
that does not consume the bulk of the development time when writing a
multi-threaded application. The challenge in writing a multi-threaded
application lies in making sure that in a chaotic, unpredictable, real-world
runtime environment threads act in an orderly, well-known manner,
avoiding such nasty conditions as deadlock and data corruption caused
by race conditions. The example in Figure 5.1 showed one Windows
mechanism for coordinating the actions of multiple threadsevents. This
section will look at the different object types Microsoft provides for
sharing data among threads.
Microsoft defines several different types of synchronization objects as
part of the Win32 API. These include events, semaphores, mutexes, and
critical sections. In addition, the Wait methods allow the developer to
wait on thread and process handles, which may be used to wait for
thread and process termination. Finally, atomic access to variables and
linked lists can be achieved through the use of interlocked functions.
Before we discuss the different data structures provided by Windows,
lets review a few of the basic concepts that are used to synchronize
concurrent access requests to shared resources. The critical section is
the block of code that can only be accessed by a certain number of
threads at a single time. In most cases, only one thread may be executing
in a critical section at one time. A semaphore is a data structure that
limits access of a particular critical section to a certain number of
threads. A mutex is a special case of a semaphore that grants exclusive
access of the critical section to only one thread. With these basic

5
Microsoft defines an additional function for signaling events: PulseEvent(). PulseEvent()
combines the functionality of SetEvent() with ResetEvent(). It is not covered in this text, other
than in this footnote, as Microsofts documentation indicates that the function is unreliable and
should not be used.
88 Multi-Core Programming

definitions in hand, we are now in a position to examine how Microsoft

implements these constructs. Generally speaking, the implementation of
these concepts is straightforward in Windows.
A semaphore object is created using the Windows
CreateSemaphore() call:
HANDLE CreateSemaphore(
LPSECURITY_ATTRIBUTES lpSemaphoreAttributes,
LONG lInitialCount,
LONG lMaximumCount,
LPCTSTR lpName );

Like the previous thread functions, CreateSemaphore() allows the

programmer to set certain security attributes that determine whether or
not the handle will be inherited by child processes. The maximum
number of threads that may be accessing the critical section protected by
the semaphore is specified by the lMaximumCount parameter. A name, as
pointed to by the lpName parameter, can be given to the semaphore.
This name can then be used by the OpenSemaphore() function to get a
handle to an already created semaphore:
HANDLE OpenSemaphore( DWORD dwDesiredAccess,
BOOL bInheritHandle,
LPCTSTR lpName );

It should be noted that by specifying a name in the lpName

parameter, the program will create a system-wide semaphore that is
available and visible to all processes. Specifying NULL for this parameter
creates a local semaphore. Unless inter-process communication is
needed, it is recommended that programs use local semaphores.
Once a semaphore is created, the developer can wait on the semaphore
using WaitForSingleObject(). WaitForSingleObject() will wait on
the semaphore handle until the thread is able to acquire the semaphore, the
specified timeout has expired, or an error occurred with the call to
WaitForSingleObject(). In the case that the thread is allowed to enter
the critical section, the semaphore count is decreased by 1 until it reaches 0.
At that point, the semaphore enters the non-signaled state and no other
threads are allowed into the critical section until one of the threads exits the
critical section by calling ReleaseSemaphore():
BOOL ReleaseSemaphore( HANDLE hSemaphore,
LONG lReleaseCount,
LPLONG lpPreviousCount );
 Chapter 5: Threading APIs 89

ReleaseSemaphore() will increment the semaphores object count

6
by the increment value specified in lReleaseCount. An example of
using semaphore objects to protect a critical section of code is shown in
Listing 5.3:

1 HANDLE hSemaphore;
2 DWORD status;
3
4 // Create a binary semaphore that is unlocked
5 // We dont care about the name in this case
6 hSemaphore = CreateSemaphore(NULL, 1, 1, NULL);
7
8 // verify semaphore is valid
9 if (NULL == hSemaphore)
10 {
11 // Handle error
12 ;
13 }
14
15 ...
16
17 // We are now testing our critical section
18 status = WaitForSingleObject(hSemaphore, 0);
19
20 if (status != WAIT_OBJECT_0)
21 {
22 // cannot enter critical section handle appropriately
23 }
24 else
25 {
26 // enter critical section
27 // time to exit critical section
28 status = ReleaseSemaphore(hSemaphore, 1, NULL);
29 if (!status)
30 {
31 // release failed, recover from error here
32 }
33 }

Listing 5.3 Using a Semaphore Object to Protect a Critical Section of Code

6
If the increment value were to cause the semaphores count to exceed the maximum count, the
count will remain unchanged, and the function will return FALSE, indicating an error condition.
Always check return values for error conditions!
90 Multi-Core Programming

A mutex in Windows works in much the same way as the semaphore

object does. The programmer makes a call to CreateMutex(), which
returns a handle which may be used by WaitForSingleObject() to
determine whether or not a thread may access a critical section. When
a thread is about to leave a critical section, it makes a call to
ReleaseMutex(), which indicates that the thread is exiting the critical
section. A mutex may be named, and may be opened by calling
OpenMutex(). As in the case of semaphores, associating a name to a
mutex will create a system wide mutex. Listing 5.4 shows how to use a
mutex object:

1 HANDLE hMutex;
2 DWORD status;
3
4 // Create a mutex
5 // Note that there arent count parameters
6 // A mutex only allows a single thread to be executing
7 // in the critical section
8 // The second parameter indicates whether or not
9 // the thread that creates the mutex will automatically
10 // acquire the mutex. In our case it wont
11 // We dont care about the name in this case
12 hMutex = CreateMutex(NULL, FALSE, NULL);
13 if (NULL == hMutex) // verify mutex is valid
14 {
15 // handle error here
16 }
17
18 ...
19
20 // We are now testing our critical section
21 status = WaitForSingleObject(hMutex, 0);
22
23 if (status != WAIT_OBJECT_0)
24 {
25 // cannot enter critical section handle appropriately
26 }
27 else
28 {
29 // enter critical section
30 // do some work
31
32 ...
33
34 // time to exit critical section
35 status = ReleaseMutex(hMutex);
36 if (!status)
 Chapter 5: Threading APIs 91

37 {
38 // release failed, recover from error here
39 }
40 }

Listing 5.4 Using a Mutex Object to Protect a Critical Section of Code

Theres one important point to note with regards to both the mutex
and semaphore objects. These objects are kernel objects, and can be
used to synchronize access between process boundaries. This ability
comes at a price; in order to acquire a semaphore, a call to the kernel
must be made. As a result, acquiring a semaphore or mutex incurs
overhead, which may hurt the performance of certain applications. In
the case that the programmer wants to synchronize access to a group
of threads in a single process, the programmer may use the
CRITICAL_SECTION data structure. This object will run in user space,
and does not incur the performance penalty of transferring control to the
kernel to acquire a lock.
The semantics of using CRITICAL_SECTION objects are different
from those of mutex and semaphore objects. The CRITICAL_SECTION
API defines a number of functions that operation on CRITICAL_SECTION
objects:
void InitializeCriticalSection( LPCRITICAL_SECTION lpCS );
void InitializeCriticalSectionAndSpinCount(
LPCRITICAL_SECTION lpCS,
DWORD dwSpinCount );
void EnterCriticalSection( LPCRITICAL_SECTION lpCS );
BOOL TryEnterCriticalSection( LPCRITICAL_SECTION lpCS );
void LeaveCriticalSection( LPCRITICAL_SECTION lpCS );
DWORD SetCriticalSectionSpinCount( LPCRITICAL_SECTION lpCS,
DWORD dwSpinCount );
void DeleteCriticalSection( LPCRITICAL_SECTION lpCS );
EnterCriticalSection() blocks on a critical section object when
it is not available. The non-blocking form of this operation is
TryEnterCriticalSection().

Atomic Operations
Acquiring mutexes and other locking primitives can be very
expensive. Many modern computer architectures support special
instructions that allow programmers to quickly perform common
atomic operations without the overhead of acquiring a lock. Microsoft
supports the operations through the use of the Interlocked API.
92 Multi-Core Programming

The Interlocked functions perform atomic operations 32-bit and

64-bit variables. These functions enable the following operations:
InterlockedIncrement() atomically increments a 32-bit
variable. The 64-bit version is InterlockedIncrement64().
InterlockedDecrement() atomically decrements a 32-bit
variable. The 64-bit version is InterlockedDecrement64().
InterlockedExchange() atomically assigns one 32-bit value
to another target
variable. The 64-bit version is
InterlockedExchange64(). To exchange pointers, use the
InterlockedExchangePointer() function.
InterlockedExchangeAdd() provides an atomic version of the
C += operator. The function atomically adds a value to a target
32-bit variable and then assigns the resulting sum to the target
variable. The 64-bit version is InterlockedExchangeAdd64().
InterlockedCompareExchange() atomically compares the
destination value with a comparison value and updates the
destination if the comparison is true. The function takes
three parameters: a pointer to the Destination variable, an
Exchange value, which is a 32-bit value to assign to Destination if
the comparison is true, and Comperand, which is the value that
Destination will be compared with. If Destination is equal to
Comperand, then Destination is assigned to the value of
Exchange. If the comparison fails, then the function doesnt
do anything. The 64-bit version of this function is
InterlockedCompareExchange64(). To exchange pointers,
use InterlockedCompareExchangePointer().
In addition to providing atomic access to variables, the Interlocked
functions enable atomic access to singly linked lists. Four operations are
defined as part of this class of operations:
InitializeSListHead() initializes the linked list.
InterlockedPushEntrySList() atomically adds a node to the
front of the list.
InterlockedPopEntrySList() atomically removes a node from
the front of the list.
InterlockedFlushSList() atomically removes all nodes in
the list.
 Chapter 5: Threading APIs 93

Thread Pools
In certain applications, the developer may need to dynamically allocate a
number of threads to perform some task. The number of threads may
vary greatly, depending on variables that are completely out of the
developers control. For example, in a Web server application, there may
be times where the server is sitting idle, with no work to be done.
During other times, the server may be handling thousands of requests at
any given time. One approach to handling this scenario in software
would be dynamic thread creation. As the system starts receiving
more and more work, the programmer would create new threads to
handle incoming requests. When the system slows down, the
programmer may decide to destroy a number of the threads created
during peak load as there isnt any work to be done and the threads are
occupying valuable system resources.
A couple of problems are associated with dynamic thread creation.
First, thread creation can be an expensive operation. During peak traffic,
a Web server will spend more time creating threads than it will spend
actually responding to user requests. To overcome that limitation, the
developer may decide to create a group of threads when the application
starts. These threads would be ready to handle requests as they come in.
This certainly helps solve the overhead problem, but other problems still
remain. What is the optimal number of threads that should be created?
How can these threads be scheduled optimally based on current system
load? At the application level, most developers dont have visibility into
these parameters, and as a result, it makes sense for the operating system
to provide some support for the notion of a thread pool.
Beginning with Windows 2000, Microsoft started providing a thread
pool API that greatly reduces the amount of code that the developer
needs to write to implement a thread pool. The principal function for
using the thread pool is QueueUserWorkItem():
BOOL QueueUserWorkItem ( LPTHREAD_START_ROUTINE Function,
PVOID Context,
ULONG Flags );

The first two parameters are of the kind youve seen before in creating
Windows threads. The routine Function() is a pointer to a function that
represents the work the thread in the pool must perform. This function
must have the form:
DWORD WINAPI Function( LPVOID parameter );
94 Multi-Core Programming

The return value is the threads exit code, which can be obtained by
calling GetExitCodeThread(). The parameter argument contains a
pointer to void. This construct is a generic way of allowing a program to
pass a single parameter or a structure containing multiple parameters.
Simply cast this parameter within the Function routine to point to the
desired data type. The Flags parameter will be examined shortly.
When QueueUserWorkItem() is called for the first time, Windows
creates a thread pool. One of these threads will be assigned to Function.
When it completes, the thread is returned to the pool, where it awaits a
new assignment. Because Windows relies on this process, Function()
must not make any calls that terminate the thread. If no threads are
available when QueueUserWorkItem() is called, Windows has the
option of expanding the number of threads in the pool by creating
additional threads. The size of the thread pool is dynamic and under the
control of Windows, whose internal algorithms determine the best way
to handle the current thread workload.
If you know the work youre assigning will take a long time to
complete, you can pass WT_EXECUTELONGFUNCTION as the third
parameter in the call to QueueUserWorkItem(). This option helps the
thread pool management functions determine how to allocate threads. If
all threads are busy when a call is made with this flag set, a new thread is
automatically created.
Threads in Windows thread pools come in two types: those that
handle asynchronous I/O and those that dont. The former rely on I/O
completion ports, a Windows kernel entity that enables threads to be
associated with I/O on specific system resources. How to handle I/O
with completion ports is a complex process that is primarily the
province of server applications. A thorough discussion of I/O completion
ports may be found in Programming Applications for Microsoft
Windows (Richter 1999).
When calling QueueUserWorkItem(), you should identify which
threads are performing I/O and which ones are not by setting the
WT_EXECUTIONDEFAULT field into the QueueUserWorkItem() Flags
parameter. This tells the thread pool that the thread does not perform
asynchronous I/O and it should be managed accordingly. Threads that do
perform asynchronous I/O should use the WT_EXECUTEINIOTHREAD flag.
When using many threads and functional decomposition, consider
using the thread pool API to save some programming effort and to allow
Windows the best possible opportunities to achieve maximum
performance
 Chapter 5: Threading APIs 95

Thread Priority
All operating systems that support threads use a priority scheme to
determine how threads should be allocated time to run on a particular
core processor. This enables important work to proceed while lesser
tasks wait for processing resources to become available. Every operating
system has a different way of handling priorities. Much of the time,
priorities are of no great concern; however, every once in a while
priorities can be important to know how a particular thread will run in
the context of competing threads.
Windows uses a scheme in which threads have priorities that range
from 0 (lowest priority) to 31 (highest priority). The Windows scheduler
always schedules the highest priority threads first. This means that
higher-priority threads could hog the system causing lower-priority
threads to starveif it wasnt for priority boosts. Windows can
dynamically boost a threads priority to avoid thread starvation. Windows
automatically does this when a thread is brought to the foreground, a
window receives a message such as a mouse input, or a blocking
condition (event) is released. Priority boosts can somewhat be controlled
by the user via the following four functions:
SetProcessPriorityBoost( HANDLE hProc, BOOL disable )
SetThreadPriorityBoost( HANDLE hThread, BOOL disable )
GetProcessPriorityBoost( HANDLE hProc, PBOOL disable )
GetThreadPriorityBoost( HANDLE hThread, PBOOL disable )
All threads are created, by default, with their priority set to normal.
After creation, a threads priority is changed using this function:
BOOL SetThreadPriority( HANDLE threadHandle,
int newPriority );
The possible values for newPriority are specified in Table 5.1,
which lists the priorities in descending order. The values are self-
explanatory.

Table 5.1 Symbolic Constants for Representing the Priorities Supported by

Windows
Symbolic Constant for Thread Priority
THREAD_PRIORITY_TIME_CRITICAL
THREAD_PRIORITY_HIGHEST
THREAD_PRIORITY_ABOVE_NORMAL
THREAD_PRIORITY_NORMAL
96 Multi-Core Programming

THREAD_PRIORITY_BELOW_NORMAL
THREAD_PRIORITY_LOWEST
THREAD_PRIORITY_IDLE

The function that returns a threads priority level is:

int GetThreadPriority( HANDLE threadHandle );
This function returns a value that translates into one of the symbolic
constants in Table 5.1. Note that Windows does not identify a threads
priority level on the 0 to 31 scale nor does it allow a threads priority to
be set with such specificity. This design enables Microsoft to define new
priority levels at a future date.
You will rarely need to boost a threads priority significantly.
However, one unique situation does warrant special consideration. In
cases where precise timing is critical, the thread performing the timing
might need to boost its priority to the highest level. This measure is used
to prevent the thread from being swapped out while it is reading from
the system timer. Immediately upon getting the timing information, the
threads priority should be reset back to normal or to its original priority.
Its important to realize that while a thread is running at highest priority,
other threads are waiting, so its generally advisable not to abuse this
priority level by using it excessively.
Priority levels affect the way a thread is scheduled. On processors
with Hyper-Threading Technology (HT Technology), the two logical
processors do not interact on the basis of thread priority. Both threads
have an equal chance of using a free resource, and their instructions are
interleaved in the most efficient way possible, without any consideration
paid to thread priority. On multi-core processors, its important to
understand that one cannot rely on priority to provide mutually-exclusive
access to a critical section. On a single-core processor, a careless
developer might make the assumption that since one thread is always a
higher priority than another that it will always execute to completion
before the lower-priority thread gets a chance to execute. This is not the
case on a multi-core processor; both the high and low priority threads
may run simultaneously.

Processor Affinity
When threads are scheduled for execution, Windows chooses which
processor should run it. The policy it uses in selecting the processor is
called soft affinity. This policy suggests to the Windows scheduler that
 Chapter 5: Threading APIs 97

threads should, as much as possible, run on the processor on which they

ran previously. The idea is that this policy makes best use of data that
might remain in processor cache.
Affinity, or the preference for a thread to run on a given processor, can
be set by the developer. This is a desirable thing to do in some situations.
For example, say a program uses four threads of which two perform
considerable I/O and two perform intensive number crunching. If these
threads are running on a dual-core system with HT Technology, four logical
processors are available to the program. If both I/O threads run on the same
physical processor core and the number-crunching threads run on the other
processor core, then the processing load will be poorly balanced. The I/O
processor core will spend considerable time waiting as both threads are
reading or writing data. Meanwhile, the processor core running the number-
crunching threads will be working continuously trying to perform the
calculations.
By use of thread affinity, a developer can place one I/O thread and
one calculation thread on the same physical processor core, and likewise
on the second processor core. Now, when the I/O thread is waiting for
data to be read or written, its resources can be temporarily made
available to the calculation thread. By this means, the processor makes
best use of the delays, and the two threads help each other while
keeping both processors completely busy.
To achieve this kind of load balancing between threads and processor
cores, you need to tell Windows on which processor to run the thread.
Windows must follow this instruction. As a result, unless you know for
sure that your preference will improve performance, as in the example, it
is best not to specify affinity because doing so will interfere with the
schedulers own optimized scheduling algorithms. Thorough testing is
the best way to know with certainty the benefits of this optimization. If
you find a compelling reason to set a threads processor affinity, you do it
with the following function:
DWORD_PTR SetThreadAffinityMask (
HANDLE threadHandle,
DWORD_PTR threadAffinityMask );
The threadAffinityMask parameter is actually an unsigned integer. It
has one bit turned on for every logical processor on which the thread
can be scheduled. The first processor core, core 0, is indicated by an
affinity mask of 0x01. The second processor core, core 1, is indicated by
an affinity mask of 0x02. A thread that can run on either one of these
processors will have an affinity mask of 0x03, and so forth.
98 Multi-Core Programming

The affinity mask must be a subset of the programs own affinity

mask. Each program, or process, can specify its processor affinity. If the
thread attempts to set its affinity to a processor core not within the
processs affinity, an error ensues.
To obtain a processs affinity mask, you call:
BOOL GetProcessAffinityMask (
HANDLE processHandle,
PDWORD_PTR processAffinityMask,
PDWORD_PTR systemAffinityMask );
To call this function, you provide the process handle, and Windows will
fill in the process affinity mask and the systems affinity mask. The latter
indicates which of the systems processors are capable of handling
threads.
If need be, the processs affinity mask can be set from inside the
program by calling:
BOOL SetProcessorAffinityMask (
HANDLE processHandle,
PDWORD_PTR processAffinityMask );
The parameters to this function are used and constructed in the same
manner as those in SetThreadAffinityMask(), discussed previously.
As discussed earlier, using affinity masks to force Windows to place a
thread on a given processor core is not always a good move. In some
cases, the interference that this practice causes with the schedulers
method of running threads can create delays. As a result, affinity should
be implemented judiciously. When testing shows that it realizes
substantial benefits, it should be used. However, situations frequently
arise where tests indicate a slight-to-moderate lift in using processor
affinity. In these cases, you want to use affinity as long as it does not
disrupt the Windows scheduler. The function to communicate affinity as
a preference rather than a command is:
DWORD SetThreadIdealProcessor (
HANDLE threadHandle,
DWORD idealProcessor );
The second parameter is not an affinity mask, but an integer
that indicates the processor core. The manifest constant
MAXIMUM_PROCESSORS can be passed for this parameter. This value
tells Windows that no particular processor core is preferred. This
practice can be useful to disable affinity for a thread that previously had
set affinity to a particular processor core. The function returns the
 Chapter 5: Threading APIs 99

previous ideal processor or MAXIMUM_PROCESSORS, if there was no

previous ideal processor core.
Important questions arise here: how do you know how many
processors a system has and which ones are configured and available?
Again, the Windows API provides a solution to this problem. Listing 5.5
shows the code that extracts processor information for the runtime
system from the operating system.

// program to obtain and display basic

// information about the runtime processor
// hardware on Windows systems

#include <windows.h>
#include <stdio.h>

void main()
{
SYSTEM_INFO sysInfo;

// Function loads sysInfo structure with data

GetSystemInfo ( &sysInfo );
// Display the data
printf ( "System hardware information: \n" );
printf ( " OEM ID: %u\n", sysInfo.dwOemId );
printf ( " Number of processors: %u\n",
sysInfo.dwNumberOfProcessors );
printf ( " Processor type: %u\n",
sysInfo.dwProcessorType );
printf ( " Active processor mask: %u\n",
sysInfo.dwActiveProcessorMask );
printf ( " Page size: %u bytes\n", sysInfo.dwPageSize );
}

Listing 5.5 How to Obtain Basic Processor Data from Windows

On a system with HT Technology running Windows XP, the output

from this code is:
System hardware information:
OEM ID: 0
Number of processors: 2
Processor type: 586
Active processor mask: 3
Page size: 4096 bytes
100 Multi-Core Programming

Notice that the number of processors is the number of logical processors.

In other words, it recognizes the capabilities of HT Technology as
distinct processors.
The active processor mask indicates which processors are
configured, or available for use by Windows. Each bit is set to 1 for each
configured processor, hence the decimal value 3 when the processors
are configured.
On a dual-core system with HT Technology, the output from running
the code in Listing 5.5 is:
System hardware information:
OEM ID: 0
Number of processors: 4
Processor type: 586
Active processor mask: 15
Page size: 4096 bytes

User-level Threading with Fibers

Until this point, only kernel threads have been discussed. These are
threads that are regularly preempted by other threads the Windows
scheduler wants to run. In addition, all these threads have been kernel
objects. This means the data regarding their status is created and
maintained in the kernel by kernel processes, such as the scheduler.
Windows offers a user-level threading package called fibers. Fibers
are completely contained in user space, and use cooperative, rather than
preemptive scheduling. Fibers run until you decide to manually swap
one out.
Here is how fibers work. A single thread can be broken down into
tasks that are swapped in and out by the application. The Windows
kernel knows nothing about the fibers; the kernel simply runs threads.
The work that the threads are doing is unknown to the kernel. When
using fibers, the thread is running the work of whichever fiber the
developer has specified in the logic of the code. In other words, the
scheduling algorithm is implemented in the application. Therefore, it is
up to the developer to manage the scheduling of fibers and when they
should run in the context of the threads time slice. However, a limitation
of fibers is that at no time can any single thread ever run more than one
fiber at a time. As such, fibers are not a mechanism that enables greater
parallelism, which means that they do not gain, at least directly, a
performance benefit from HT Technology or multi-core platforms. The
primary purpose of fibers is to provide the developer with a convenient
 Chapter 5: Threading APIs 101

method of scheduling multiple tasks that are known to not need parallel
execution.
The first step in using Windows fibers is to convert the current
thread into a fiber. Once this is done, additional fibers can be added. So,
the following function is the first one to call:
PVOID ConvertThreadToFiber( PVOID parameters );
This function returns the address of the fibers internal data area, which
contains housekeeping items. This address should be saved. Later on, when
you switch fibers, the address of this data area will be needed. The sole
parameter to this function is a pointer to arguments for this fiber. It seems a bit
strange for a thread to pass arguments to itself. However, this parameter can be
retrieved from the fibers internal data area using the function:
PVOID GetFiberData();
There is no point in converting a thread into a fiber unless you plan
to run multiple fibers on the thread. So, once youve converted the
thread into a fiber, you should next create the other fibers you plan to
run. The function to do this is:
PVOID CreateFiber ( DWORD fiberStackSize,
PFIBER_START_ROUTINE fiberProc,
PVOID fiberProcParameters );
The first parameter specifies how large the stack for the fiber should
be. Normally, a value of 0 is passed here. Passing a 0 will cause
Windows to allocate two pages of storage and to limit the stack size
to the default 1 MB. The next two parameters should look familiar
from thread-creation functions you have previously seen. The first is
a pointer to a fiber function; the second is a pointer to the
parameters of that function. Note that unlike a thread function this
fiber function does not return a value. This function has the form:
VOID WINAPI fiberProc( PVOID fiberProcParameters );
An important characteristic of the fiber function is that it must not exit.
Remember that when using threads and the thread function exits, the
thread is terminated. However, with fibers, the effect is more dramatic:
the thread and all the fibers associated with it are terminated.
Again, its important to save the address returned by CreateFiber()
because it is used in the following function to switch among the fibers:
VOID SwitchToFiber( PVOID addressOfFiberEnvironment );
The sole parameter to this function is the address returned by
CreateFiber() and ConvertThreadToFiber(). Switching to a fiber is
102 Multi-Core Programming

the only way to activate a fiber. You can switch anytime you desire to.
You basically receive total control over scheduling in exchange for the
fact that only one fiber at a time can run on a thread. Only a fiber can
switch to another fiber. This explains why you must convert the original
thread into a fiber at the start of this process.
The function to delete a fiber is:
VOID DeleteFiber( PVOID addressOfFiberEnvironment );
A fiber can kill itself this way. However, when it does so, it kills the
current thread and all fibers associated with it.
A final function that is useful is
PVOID GetCurrentFiber();
which returns the address of the fiber environment of the currently
executing fiber.
Listing 5.6 shows the code for a program that creates some fibers and
has them print their identification.

// demonstration of the use of Windows fibers

#define _WIN32_WINNT 0x400

#include <stdio.h>
#include <windows.h>

#define FIBER_COUNT 10
void *fiber_context[FIBER_COUNT];

VOID WINAPI fiberProc ( void * );

void main()
{
int i;
int fibers[FIBER_COUNT];

for ( i = 0; i < FIBER_COUNT; i++ )

fibers[i] = i;

fiber_context[0] = ConvertThreadToFiber ( NULL );

for ( i = 1; i < FIBER_COUNT; i++ )

{
fiber_context[i] = CreateFiber (
0, // stack size
fiberProc, // function
&fibers[i] ); // parameter
 Chapter 5: Threading APIs 103

if ( fiber_context[i] != NULL )
printf ( "fiber %d created\n", i );
}

for ( i = 1; i < FIBER_COUNT; i++ )

{
if ( fiber_context[i] != NULL )
SwitchToFiber ( fiber_context[i] );
}
}

VOID WINAPI fiberProc ( void *fiber_nbr )

{
int nbr;

nbr = *( (int*) fiber_nbr );

printf ( "Hello from fiber %d\n", nbr );

// now switch back to the fiber of the main line

SwitchToFiber ( fiber_context[0] );
}

Listing 5.6 Program to Create Fibers that Print an Identifying Message to the Console

Notice the #defined manifest constant at the very start of the listing.
Fibers were introduced in Windows NT 4.0. The value of 0x400 in:
#define _WIN32_WINNT 0x400
tells the compiler to include features in windows.h that appeared in
Microsoft Windows NT 4.0 and later; hence, it includes support for
function calls used by the fiber APIs. Failing to include the constant will
result in compilation errors. The output from this program is:
fiber 1 created
fiber 2 created
fiber 3 created
fiber 4 created
fiber 5 created
fiber 6 created
fiber 7 created
fiber 8 created
fiber 9 created
Hello from fiber 1
Hello from fiber 2
Hello from fiber 3
Hello from fiber 4
Hello from fiber 5
Hello from fiber 6
104 Multi-Core Programming

Hello from fiber 7

Hello from fiber 8
Hello from fiber 9
Notice how the fibers print out their messages in a nice sequential order.
This is, of course, attributable to the fact that only one fiber can run at a
time and that the main fiber schedules which one runs when.
It might seem that fibers are not valuable in the context of parallel
programming because they only run one at a time. However, they have
considerable value in certain specific applications. Suppose for example
that your application receives numerous requests that you want to
handle individually. You could create a thread for each request, but this
approach has drawbacks. Threads are expensive to start up and you
dont have control over their execution. The former problem can be
handled with thread pools, but because pools are generally used for all
thread tasks a program must perform, the execution schedule is even
more uncertain. In addition, if you have to expand your thread pool, the
creation of threads is a kernel call and, hence, expensive.
Fibers, however, are cheap to create because they do not involve
user-space to kernel transitions. In fact, several of the fiber functions are
implemented as inline functions. Fibers are an efficient mechanism any
time you have multiple tasks to perform and are content doing one at a
time. Notice that you can emulate parallel processing by swapping fibers
as needed. This way, if a fiber is about to wait on a slow process, you can
swap in another fiber and give it as much time as you deem necessary. In
some situations where the behavior and parallelism of the application is
well known, fibers may even provide a performance benefit due to the
reduction of synchronization and task-switching overhead.

Compiling and Linking Multi-threaded Applications in Windows

In order to simplify the building of multi-threaded applications for
Windows platforms, Microsoft has built in support for multithreading
into the Microsoft compiler. The first step in compiling a multi-threaded
application is to make sure that the compiler is using the correct runtime
libraries. Microsoft provides several different implementations of the
standard C runtime library. These versions vary based on three different
criteria: linkage (static or dynamic), whether or not the library is re-
7
entrant, and whether or not the library includes debugging symbols. The

7
One library is designed for single-threaded applications; it is not re-entrant. The other library is
designed for multi-threaded applications; it is re-entrant.
 Chapter 5: Threading APIs 105

compiler switches that are used in multi-threaded applications are shown

in Table 5.2. These compiler switches can be specified on the command
line or by selecting the appropriate option in the code generation tab
of the projects properties within Visual Studio, as shown in Figure 5.1.

Table 5.2 Command-line Switches for Compiling Multi-threaded Code in

Windows.
Switch Meaning
/MT Multithreaded with static linkage
/MTd Multithreaded with debugging enabled
/MD Multi-threaded dynamic link library (DLL)
/MDd Multi-threaded DLL with debugging enabled

Figure 5.1 Enabling Multi-threading in Microsoft Visual Studio

In addition, the _MT preprocessor value should be defined. This is the

agreed convention in Windows for testing code for multithreading. It is
set by the /D "_MT" command-line switch.
106 Multi-Core Programming

The appropriate libraries are generally brought in by setting the

correct flag from Table 5.2. However, it is important to know more about
which libraries are used. This knowledge will help in the case that you
need to troubleshoot compiler issues or you want to write custom
makefiles. Table 5.3 provides the list of multi-threaded libraries that are
used when linking in the Microsoft C runtime library.
Table 5.3 Microsofts Multi-threaded Libraries for C Programs
Program Type C Runtime Library
Multithreaded LIBCMT.lib
Multithreaded with debugging LIBCMTD.lib
Multi-threaded DLL MSVCRT.lib (import library for MSVCRT.dll)
Multi-threaded DLL with MSVCRTD.lib (import library for MSVCRTD.dll)
debugging

The libraries for the C++ runtime are different, as shown in Table 5.4

Table 5.4 Microsofts Multi-threaded Libraries for C++ Programs

Program Type C++ Runtime library
Multithreaded LIBCPMT.lib
Multithreaded with debugging LIBCPMTD.lib
Multithreaded DLL MSVCPRT.lib (also uses MSVCRT.dll)
Multithreaded DLL with debugging MSVCPRTD.lib (also uses MSVCRTD.dll)

A word of caution is in order here. Developers should not intermix

the use of static and dynamic libraries in a single process space. Multiple
copies of the C runtime library within a single process space will result in
8
unstable behavior. The Microsoft linker will prevent this error in the
case where the developer tries to link both the static and dynamic
versions of the runtime library into a single .exe. However, in the case of
a .dll that is used by an .exe, it is the programmers responsibility to
guarantee that the .exe and .dll are built with the same version of the C
runtime.

8
Note that this is the second time that using the standard C library routines has introduced an
additional level of complexity to using threads in Windows; the first was when using the
CreateThread() call. In general, Microsoft encourages the use of the Win32 API over the standard
C library, for instance, CreateFile() instead of fopen(). Using the Win32 API exclusively will
simplify writing Windows-based multi-threaded applications.
 Chapter 5: Threading APIs 107

Threading APIs for Microsoft .NET Framework

In 2002, Microsoft introduced a new execution environment, called the
Common Language Runtime (CLR), which executes programs in the form
of intermediate code, rather than native binaries. This code is
conceptually similar to Java bytecodes in that the CLR executes them
within a carefully defined virtual machine. And like Java, it frequently
uses just-in-time compilation to convert the intermediate code into binary
routines on the fly.
The goal of this design was to provide a common execution
environment for all the major languages Microsoft supported at the time,
notably C++ and Visual Basic, and for a new language, named C#, that
Microsoft unveiled with the launch of the CLR and associated
technologies. Those technologies included an extensive set of APIs
commonly referred to as the .NET Framework. Together, the framework,
the CLR, and the tools needed for execution, are referred to as .NET.
The .NET Framework has extensive support for threads. The API
provides a substantial subset of the functionality of the Windows API. It
does not include some features, such as fibers, but most thread
management functions and concepts are present. These are implemented
in the .NET Framework class entitled Thread. This section discusses how
to use Thread in .NET. The examples are implemented in C#. If you want
to use another .NET language, such as C++, Visual Basic .NET, or JScript,
consult Microsofts .NET Framework Class Library Reference, which
provides examples of the APIs in all four languages.

Creating Threads
On the whole, .NET APIs tend to be somewhat leaner than their Win32
counterparts. This is especially visible in the call for creating a new
thread:
using System.Threading;
. . .
Thread t = new Thread( new ThreadStart( ThreadFunc ));

The call to ThreadStart() constructs a new thread. The parameter is a

delegate called ThreadFunc. In C#, a delegate is the equivalent of an
address of a function in C. Its a manner of identifying a function
or method without actually invoking it. As with Win32, when
ThreadFunc() ends, the thread terminates.
108 Multi-Core Programming

Listing 5.7 illustrates a simple creation of a thread and the call to the
ThreadFunc.

1 using System;
2 using System.Threading;
3
4 public class ThreadDemo1
5 {
6 public static void ThreadFunc()
7 {
8 for ( int i = 0; i < 3; i++ )
9 Console.WriteLine(
10 "Hello #{0} from ThreadFunc", i );
11 Thread.Sleep( 10000 );
12 }
13
14 // The main entry point for the application.
15 public static void Main()
16 {
17 Thread t =
18 new Thread( new ThreadStart( ThreadFunc ));
19 t.Start();
20 Thread.Sleep( 40 );
21
22 for ( int j = 0; j < 4; j++ )
23 {
24 Console.WriteLine( "Hello from Main Thread" );
25 Thread.Sleep( 0 );
26 }
27 }
28 }

Listing 5.7 Simple Program Showing Thread Creation

A new thread is created using ThreadStart() on lines 1718. The

ThreadFunc is defined in lines 612. An important difference between
the Win32 and .NET APIs appears here. In Win32, the creation of a
thread results in a request to the scheduler to execute it. No additional
step is necessary. In .NET, this is not so. The thread must be explicitly
started by the start() method, as shown on line 19. This method
changes the status of the thread to Runnable inside .NET, which makes
it eligible for processing by the scheduler. On a single processor system,
however, .NET will not start the thread until the main thread has
completed or is suspended. This can cause obvious problems, so it is
common to suspend the current, principal thread to enable created
 Chapter 5: Threading APIs 109

threads to begin execution. This is done on line 20 by the call to the

Sleep() method. The sole argument to this method is a 32-bit integer
that indicates the number of milliseconds to sleep before being
awakened, that is, reactivated. As with nearly all things related to threads,
nothing is guaranteed. So, the number of milliseconds is simply a request
to the operating system. The actual number of milliseconds in the
sleeping state might vary considerably from the requested amount.
Because Microsoft is aware that Sleep() must be called to start up
waiting threads, it provides a special value of 0 to indicate that the
principal thread should be suspended only long enough to start
execution of the any waiting threads. On the other end of the scale, the
special value Infinite tells the scheduler to suspend the thread
indefinitely.
As Listing 5.7 illustrates, the method for creating a function in C# is
much more concise than the corresponding function call in C on Win32.
This concision comes at a cost, however. The delegate called when a
thread is started accepts no parameters, so it is difficult to pass thread-
specific information to the delegate. To pass data without making it
accessible globally is an advanced topic, which is beyond the scope of this
book. (Abrams 2004), however, explains how to do this.

Thread Priority and Other Properties

As discussed in the previous sections on the Win32 API, thread priority
determines when the scheduler will schedule execution. Threads with
higher priority can grab the processor for themselves and even starve
lower priority threads, so managing priority in special circumstances can
be beneficial. For the most part, you should accept the default priority
that .NET assigns to your threads. However, if you need to change it, you
simply specify it as an assignment to the thread.Priority property.
For example,
Thread nThread = new Thread ( ThreadFunc );
nthread.Priority = AboveNormal;
nThread.Start ();
The .NET framework supports five levels of thread priority:
Highest
AboveNormal
Normal (the default level)
BelowNormal
Lowest
110 Multi-Core Programming

Each of these levels represents a group of fine-grained levels, which are

accessible through Win32. When the thread is started, the scheduler
determines where in the group of levels to place the thread. For
example, if you specify BelowNormal, the scheduler chooses the sub-
level within BelowNormal to place your thread. You have no control
over this decision.
Another useful property .NET provides enables you to name a thread.
This option has several benefits, the most important of which is in
debugging, where the name makes it much easier for developers to
identify which thread is being followed. To name a thread, simply
initialize its .name property. For example:
Thread nThread = new Thread ( ThreadFunc );
nthread.Name = "new_thread1";
nThread.Start ();
Having examined thread creation, lets examine how threads are
terminated and how their activities can be suspended and restarted.

Managing Threads
The simplest and safest way to terminate a thread is to exit it. Doing so,
permits the CLR to perform the necessary clean up without any difficulty.
At times, however, its necessary to terminate some other thread. As part
of the .NET threading API, an Abort() method is supplied for this
purpose. A call to Abort() generates a ThreadAbortException, where
any code should go to handle an abort signal in the middle of an
unfinished operation. In addition, the call will execute any code in the
aborted threads finally statement. Listing 5.8 shows the calls.

1 using System;
2 using System.Threading;
3
4 public class ThreadAbortExample
5 {
6 public static void Thread2()
7 {
8 try
9 {
10 Console.WriteLine( "starting t2" );
11 Thread.Sleep( 500 );
12 Console.WriteLine( "finishing t2" );
13 }
14 catch( ThreadAbortException e )
 Chapter 5: Threading APIs 111

15 {
16 Console.WriteLine( "in t2\'s catch block");
17 }
18 finally
19 {
20 Console.WriteLine( "in t2\'s finally" );
21 }
22 }
23
24 public static void Main()
25 {
26 Thread t = new Thread( new ThreadStart(Thread2) );
27 Console.WriteLine( "starting main thread" );
28 t.Start();
29 Thread.Sleep( 500 );
30 t.Abort();
31 t.Join();
32 Console.WriteLine( "main thread finished.\n" +
33 "Press <Enter> to exit" );
34 Console.ReadLine();
35 }
36 }

Listing 5.8 How to Abort a Thread in .NET

Lines 2628 start a second thread, t, as illustrated previously. After a

brief pause (line 29), t is aborted. The effect of this is to call the catch
code in Thread2s exception handler, followed by the code in the
finally clause. The output of this program is:
starting main thread
starting t2
finishing t2
in t2's catch block
in t2's finally
main thread finished.
Press <Enter> to exit
Calling Abort() results in a more complicated set of actions. The
first is that the thread being aborted can thwart the action by calling
the System.Threading.Thread.ResertAbort method inside the
exception handler.
Another factor should be borne in mind: processing in the finally
block might take a substantial amount of time. Because the thread will
stay alive throughout the processing of the finally code, the thread might
not abort for a substantial period of time, sometimes not until the
112 Multi-Core Programming

program itself is finished. Because of these two possibilities of a thread

persisting past the initial Abort(), it might be necessary to verify that a
specific thread has indeed been aborted. The way to do this is to use the
join method, which is discussed next.

Waiting on a Thread
Threads often need to wait on each other. This concept is presented in
the Win32 APIs as waiting on an event. The .NET Framework borrows
the model used by Pthreadsthe API employed in Linux and several
versions of UNIX. There the concept is known as joining a thread, and
it simply means waiting for that thread to finish. Line 31 of Listing 5.8
shows how this method is called. In that program, the main thread
creates and aborts Thread2 then joins it. This is the preferred way of
knowing with certainty that a given thread has aborted.
It is important to note that the thread calling Join() blocks until the
joined thread exits. In some circumstances, this might not be desirable
because in such cases, Join() can be called with a 32-bit integer
parameter, which indicates the maximum number of milliseconds to wait
for the joined thread to complete. When called, this way, Join() returns
the Boolean value true if the thread terminated, and false if the thread
did not terminate and the return occurred because the maximum wait
expired.

Suspending and Resuming Threads

Earlier in this section, we described the use of the Sleep() method to
suspend a thread for a time-delimited duration. Examples of its use
appear in lines 11 and 20 of Listing 5.8.
There are times, however, when its desirable to suspend a thread for
an indefinite period of time and then resume it at a later point. The pair
of methods to do this are Suspend() and Resume(). Neither of these
methods takes an argument or returns a value.

Thread Pools
The creation of a new thread is an expensive operation. A lot of system-
level activity is generated to create the new thread of execution, create
thread-local storage, and set up the system structures to manage the
thread. As a result of this overhead, conservation of created threads is a
recommended practice. The effect on performance, especially on slower
machines, can be compelling.
 Chapter 5: Threading APIs 113

However, managing the task of managing multiple threads, some of

which might be dormant, and assigning new chunks of work to the
threads as the work becomes available, leads to a significant amount of
code complexity. In response, the .NET Framework provides a resource
called thread pools.
A pool of threads is initially created as well as a work queue of sorts.
As new chunks of work become available they are queued up for the
thread pool, which executes them as threads are available. The process
of managing the work queue, waking threads, and assigning them the
work is all handled by the .NET Framework. This is a valuable resource. If
your program uses more than a few threads, thread pools should be
considered as an option.
.NETs ThreadPool class does most of the work of thread pools. The
pool is created the first time work is queued for ita remarkably high-
level operation. The exact number of threads created in the pool is
dynamic and determined by .NET when the pool is created. However,
.NET enforces a maximum of 25 threads per hardware processor. The
ThreadPool methods GetMinThreads() and SetMinThreads() can be
used to inquire and enforce a minimum number of threads. A
corresponding GetMaxThreads() method informs you of the maximum,
but there is no method that enables you to increase this value.
Listing 5.9 shows a simple example of a thread pool in use.

1 using System;
2 using System.Threading;
3 public class ThreadPoolExample
4 {
5 public static void Main()
6 {
7 // Queue a piece of work
8 ThreadPool.QueueUserWorkItem(
9 new WaitCallback( WorkToDo ));
10
11 Console.WriteLine( "Greetings from Main()" );
12 Thread.Sleep( 1000 );
13
14 Console.WriteLine( "Main thread exiting...\n" +
15 "Press <enter> to close" );
16 Console.ReadLine();
17 }
18
19 // This thread procedure performs the task.
20 static void WorkToDo( Object dataItems )
114 Multi-Core Programming

21 {
22 Console.WriteLine("Greetings from thread pool");
23 }
24 }

Listing 5.9 Simple Example of Using a Thread Pool in .NET

The thread pool is created by the first call to the work queue, which
occurs in line 8. As in thread creation, it is passed a delegate; which, in this
case points to the method defined in lines 2023. As can be seen from the
signature on line 20, there is an overloaded version, which permits an
object to be passed to the work procedure. Frequently, this data object
contains state information about the status of the application when the call
work was queued, but it can, in fact, contain any data object.
Notice the call to Sleep() on line 12. It is necessary for successful
completion of this program. Without this statement, the program could
exit without the work queue ever having completed its work. Because
the work can be assigned to any available thread, the main thread has no
way to join any of the pools threads, so it has no mechanism for waiting
until they complete. Of course, the threads in the pool can modify a data
item to indicate activity, but that is a not a .NET-specific solution.
The output from this program is:
Greetings from Main()
Greetings from thread pool.
Main thread exiting...
Press <enter> to close
In addition to being work engines that consume queued work items,
thread pools are effective means of assigning threads to wait on specific
events, such as waiting on network traffic and other asynchronous
events. The .NET Framework provides several methods for waiting. They
require registering a call-back function that is invoked when the waited-
for event occurs. One of the basic methods for registering a call-back and
waiting is RegisterWaitForSingleObject(), which enables you to
also specify a maximum wait period. The call-back function is called if
the event occurs or the wait period expires. Listing 5.10, which is
adapted from a Microsoft example, shows the necessary code.

1 using System;
2 using System.Threading;
3 // TaskInfo contains data that is passed to the callback
4 // method.
5 public class TaskInfo
 Chapter 5: Threading APIs 115

6 {
7 public RegisteredWaitHandle Handle = null;
8 public string OtherInfo = "default";
9 }
10
11 public class Example
12 {
13 public static void Main( string[] args )
14 {
15
16 AutoResetEvent ev = new AutoResetEvent( false );
17
18 TaskInfo ti = new TaskInfo();
19 ti.OtherInfo = "First task";
20 ti.Handle =
21 ThreadPool.RegisterWaitForSingleObject(
22 ev,
23 new WaitOrTimerCallback( WaitProc ),
24 ti,
25 1000,
26 false );
27
28 // The main thread waits three seconds,
29 // to demonstrate the time-outs on the queued
30 // thread, and then signals.
31
32 Thread.Sleep( 3100 );
33 Console.WriteLine( "Main thread signals." );
34 ev.Set();
35
36 Thread.Sleep( 1000 );
37 Console.WriteLine( "Press <enter> to close." );
38 Console.ReadLine();
39 }
40
41 // The callback method executes when the registered
42 // wait times-out, or when the WaitHandle (in this
43 // case, AutoResetEvent) is signaled.
44
45 public static void WaitProc( object passedData,
46 bool timedOut )
47 {
48 TaskInfo ti = (TaskInfo) passedData;
49
50 string cause = "TIMED OUT";
51 if ( !timedOut )
52 {
53 cause = "SIGNALED";
54 if ( ti.Handle != null )
116 Multi-Core Programming

55 ti.Handle.Unregister( null );
56 }
57
58 Console.WriteLine(
59 "WaitProc({0}) on thread {1}; cause={2}",
60 ti.OtherInfo,
61 Thread.CurrentThread.GetHashCode().ToString(),
62 cause
63 );
64 }
65 }

Listing 5.10 Using Callbacks in Thread Pools to Wait on Events

There is a lot going on in this listing. The method that is registered

for the callback is WaitProc, defined in lines 45-64. As can be seen, it
takes two parameters: the first is an object containing data to be passed
to the method, the second is a Boolean to indicate whether the call was
generated by a time out (true) signal from the waited for event (false).
The passed data object is cast to an object of type TaskInfo, which is
defined in lines 39. The handle property of TaskInfo is returned from
the call to RegisterWaitForSingleObject() on lines 21-26. The
parameters to this method are complex. ev is a handle for the event
being waited forwell come back to this shortly. The second parameter
is the delegate for the callback function; as can be seen, that delegate
must be of type WaitOrTimerCallback.
The third parameter is the data object to pass to the callback function.
The fourth parameter is the number of milliseconds to wait. And the fifth is
a Boolean that indicates whether the event should stop waiting after the
delegate has been called (true) or keep wait anew (false).
As soon as the handle is created, the wait begins. The Sleep
statement on line 32 allows this wait to expire several times. Because
of the parameter on line 26, the wait renews and expires several times,
each time calling the method pointed to by the delegate. Line 34 actually
triggers the event via a direct call. The callback function is able to
distinguish which event triggered the call by the timedOut parameter
discussed previously. As a result, running this code generates the
following output:
WaitProc(First task) executes on thread 4; cause=TIMED OUT.
WaitProc(First task) executes on thread 4; cause=TIMED OUT.
WaitProc(First task) executes on thread 6; cause=TIMED OUT.
Main thread signals.
WaitProc(First task) executes on thread 6; cause=SIGNALED.
Press <enter> to close.
 Chapter 5: Threading APIs 117

The number of the thread on which the task is executed will vary
from system to system. As can be seen, while the main thread is waiting,
the 1-second duration expires three times, as expected. Then, the
callback function is called one more time when the signal is sent.
The .NET Framework enables threads to start up based on more than
a single event. The WaitHandle.WaitAll() and WaitHandle.Wait-
Any()methods fire when all events in an array have been signaled, or if
any one event in an array is signaled, respectively. Events themselves do
not need to be automatic as in Listing 5.10; they can also be manual by
using ManualResetEvent(). The difference is that an automatic reset
will issue the signal and then reset itself so that it is not in the signaled
state, whereas a manual reset event persists in the signaled state until it is
manually reset. The choice between them depends entirely on the
applications needs.
As this section has illustrated, thread pools are a very useful
mechanism that enables sophisticated threading to be implemented
conveniently in many applications. The range of options regarding events
and the characteristics of signals give thread pools considerable
flexibility.

Thread Synchronization
The mechanisms for synchronizing thread actions in .NET are similar to
those found in all other threading APIs, such as Win32 and Pthreads.
They include capabilities for mutual exclusion and for atomic actions
on specific variables. By and large, .NET maintains the simplicity of
expression seen in the previous examples. No synchronization is simpler,
in fact, than use of the lock keyword in C#.
The usual way to use lock is to place it in front of a block of code
delimited by braces. Then, that block can be executed by only one
thread at a time. For example:
lock(this)
{
shared_var = other_shared_var + 1;
other_shared_var = 0;
}
The C# lock statement makes several calls to the .NET Framework.
The previous example is syntactically equivalent to the following snippet:
Monitor.Enter( this )
try
{
shared_var = other_shared_var + 1;
118 Multi-Core Programming

other_shared_var = 0;
}
finally
{
Monitor.Exit( this )
}
Monitor is a class that enforces mutual exclusion and locking in .NET.
When used as in the previous example, Monitor.Enter() locks a code
block. In this respect, it is similar to critical sections in the Win32 API.
Monitor can also be used to lock a data structure by passing that data
structure as a parameter to the Monitor.Enter() call. Monitor.Exit()
releases the lock. If Monitor.Enter() was called with an object,
Monitor.Exit() should be called with the same object to release the
lock. When Monitor.Enter() is called, the .NET Framework sets up two
queues: one containing references to threads waiting to obtain the lock
once its released, and another queue containing references to threads that
want to be signaled that the lock is available. When Monitor.Exit() is
called, the next thread in the first queue gets the lock.
Monitors have unusual aspects. For example, the Monitor.Wait()
method enables a thread to temporarily give up a lock to another thread and
then reclaim it. A system of signals called pulses are used to notify the
original thread that the lock has been released.
As you have learned, mutexes are a similar mechanism for providing
mutual exclusion to resources. Mutexes differ from monitors in that they
can be used with wait handles, as shown in the following example. They
also can be locked multiple times. In such a case, they must be unlocked
the same number of times before the lock is actually released.
To use a mutex, one must be created. Then a call to WaitOne is
issued to grab the lock as soon as it becomes available, if its not already
available. Once the lock is no longer needed, it is made available with the
ReleaseMutex method.
private static Mutex mutx = new Mutex();
. . .

private static void UseResource()

{
// Wait to enter the locked code.
mutx.WaitOne();

Console.WriteLine( " in the locked code " );

Thread.Sleep( 100 );
 Chapter 5: Threading APIs 119

Console.WriteLine( " leaving locked code ",

// Release the Mutex.

mutx.ReleaseMutex();
}
This is a simple, effective locking mechanism that is comparable to
counterparts in Win32 and Pthreads.

Atomic Actions
Actions are atomic if they can only be performed as a single indivisible
act. The term is commonly used in database operations to refer to a series
of steps that must all be completed. If any of them cant be completed, all
steps so far completed are rolled back, so that at no time does the
database record a partial series. Its all or nothing. Threads present similar
problems. Consider what happens if a thread is suspended while it is
updating the values of an important variable. Suddenly, the application or
9
the system can be left in a degenerate or corrupted state. One solution is
the Interlocked class. Although not discussed in the Win32 portion of
this chapter, the Win32 API does have corresponding APIs.
The three most common methods of the Interlocked class are:
Decrement, Increment, and Exchange. These are all simple methods to
use and should be used anytime a variable shared between threads is
being modified.
int intCounter = 0;
. . .

// set the value of intCounter to 6

Interlocked.Exchange( ref usingResource, 6 );

// Drop value to 5
Interlocked.Decrement( ref intCounter );

//Raise it back to 6
Interlocked.Increment( ref intCounter );
Several aspects are worthy of note. Firstly, the Interlocked class
uses references to the variables to be modified, not the variables
themselves; so make sure to include the ref keyword, as in the

9
It might come as a surprise to some readers that incrementing or decrementing a variable is not
inherently an indivisible action. It takes three instructions: the variable is copied into a register in
the processor core by a process called loading, incremented, and then copied from the register
back to the variables location in memory.
120 Multi-Core Programming

previous examples. These references are the equivalent of addresses

or pointers from C and C++. Secondly, the Exchange method is not
really used for exchanging values but initializing one value to another.
In the example, the value of 0 to which intCounter was initialized is
exchanged with 6, leaving 6 as the new value of intCounter.
As can be seen from this overview, the .NET Frameworks set of
threading APIs is more succinct than its counterparts in Win32. It
also includes higher-level capabilities, such as advanced thread pool
manage-ment functions. The upshot is that programming threads in C#
on .NET tends to be easier and overall more productive than at the lower
levels of Win32.

POSIX Threads
POSIX threads, or Pthreads, is a portable threading library designed with
the intent of providing a consistent programming interface across multiple
operating system platforms. Pthreads is now the standard threading
interface for Linux and is also widely used on most UNIX platforms. An
open-source version for Windows, called pthreads-win32, is available as
well. For more information on pthreads-win32, refer to References. If you
want to work in C and need a portable threading API that provides more
direct control than OpenMP, pthreads is a good choice.
Most core Pthreads functions focus on thread creation and destruction,
synchronization, plus a few miscellaneous functions. Capabilities like
thread priorities are not a part of the core pthreads library, but instead are
a part of the optional capabilities that are vendor specific.

Creating Threads
The POSIX threads call to create a thread is pthread_create():
pthread_create (
&a_thread, // thread ID goes here
NULL, // thread attributes (NULL = none)
PrintThreads, // function name
(void *) msg ); // parameter
As in Windows, the third parameter represents a pointer to the function
called by the launched thread, while the fourth parameter is a pointer to a
void, which is used to pass arguments to the called function.
Listing 5.11 illustrates the usage of pthread_create() to create a
thread.
 Chapter 5: Threading APIs 121

1 #include <stdio.h>
2 #include <stdlib.h>
3 #include <pthread.h>
4
5 void *PrintThreads ( void * );
6
7 #define NUM_THREADS 9
8
9 int main()
10 {
11 int i, ret;
12 pthread_t a_thread;
13
14 int thdNum [NUM_THREADS]; //thread numbers go here
15
16 for ( i = 0; i < NUM_THREADS; i++ )
17 thdNum[i] = i;
18
19 for ( i = 0; i < NUM_THREADS; i++ )
20 {
21 ret = pthread_create (
22 &a_thread,
23 NULL,
24 PrintThreads,
25 (void *) &thdNum[i] );
26
27 if ( ret == 0 )
28 printf ( "Thread launched successfully\n" );
29 }
30
31 printf ( "Press any key to exit..." );
32 i = getchar();
33 return ( 0 );
34 }
35
36 // Make the threads print out their thread number.
37
38 void *PrintThreads ( void *num )
39 {
40 int i;
41
42 for ( i = 0; i < 3; i++ )
43 printf ( "Thread number is %d\n",
44 *((int*)num));
45
46 return ( NULL );
47 }

Listing 5.11 Creating and Using Threads with Pthreads

122 Multi-Core Programming

The main loop does print a notice when a thread is launched

successfully, and then has the thread print its thread number three times.
Note that the output from this program does not show an orderly
sequence of print statements from each individual thread. Instead, the
print statements are printed out in a mixed order.

Managing Threads
When a thread is created under Pthreads, developers have the option of
indicating the nature of that threads interaction with other threads. For
example,
pthread_detach( pthread_t thread_to_detach );
can be used to detach the thread from the other threads when it has no
need to interact with them. This option asserts that no other thread will
interact with this thread, and that the operating system is free to use
this information in managing the thread. The operating system uses this
information particularly at thread exit, when it knows that no return value
needs to be passed back to some other thread.
The complementary function,
pthread_join( pthread_t thread, void **ret_val );
tells the operating system to block the calling thread until the specified
thread exits. Attaching to a thread in this way is called joining, just as we
saw in the section on .NET threads. The function takes two parameters:
the pthread_t identifier of the thread being joined, and a pointer to a
pointer to void where the threads return value should be placed. If the
thread does not return a value, NULL can be passed as the second
parameter.
To wait on multiple threads, simply join all those threads. Listing 5.12
shows how this is done.

int main()
{
int i, ret;

pthread_t thdHandle [NUM_THREADS]; //thread identifiers

int thdNum [NUM_THREADS]; //thread numbers go here

for ( i = 0; i < NUM_THREADS; i++ )

thdNum[i] = i;

for ( i = 0; i < NUM_THREADS; i++ )

 Chapter 5: Threading APIs 123

{
ret = pthread_create (
&thdHandle[i],
NULL,
PrintThreads,
(void *) &thdNum[i] );

if ( ret == 0 )
printf ( "Thread launched successfully\n" );
}

// join all the threads and wait...

for ( i = 0; i < NUM_THREADS; i++ )
pthread_join ( thdHandle[i], NULL );

printf ( "Press any key to exit..." );

i = getchar();
return ( 0 );
}

Listing 5.12 Coordinating Thread Execution with pthread_join

One caveat should be noted: two threads cannot join the same
thread. Once a thread has been joined, no other threads can join it. To
have two or more threads wait on a threads execution, other devices
such as those presented in the section on signaling can be used.

Thread Synchronization
The Pthreads library has mutexes that function similarly to those in
Win32 and .NET. Terminology and coding syntax, predictably, are
different; as are some details of implementation.
Whereas Windows refers to mutexes as being signaled, that is,
available or unlocked, Pthreads refers to mutexes by the more intuitive
terms locked and unlocked. Obviously, when a mutex is locked, the code
its protecting is not accessible. The syntax of the Pthreads API calls
follows this nomenclature:
pthread_mutex_lock( &aMutex );
. . . code to be protected goes here . . .
pthread_mutex_unlock( &aMutex );
The sole parameter to both functions is the address of a previously
declared mutex object:
pthread_mutex_t aMutex = PTHREAD_MUTEX_INITIALIZER;
124 Multi-Core Programming

PTHREAD_MUTEX_INITIALIZER is a macro that initializes the opaque

data type pthread_mutex_t. Certain uncommon forms of mutex
creation can use other macros; however, the vast majority of the time
you create a mutex, this is the initialization code youll want.
Using a mutex, the code for the PrintThreads() routine of
Listing 5.11 (lines 3847) would now looks like this:
void *PrintThreads( void *num )
{
int i;

pthread_mutex_lock( &testMutex );

for ( i = 0; i < 3; i++ )

printf ( "Thread number is %d\n",
*((int*)num));

pthread_mutex_unlock( &testMutex );

return ( NULL );
}
Earlier in the program, at the global level, the following definition
appeared:
pthread_mutex_t testMutex = PTHREAD_MUTEX_INITIALIZER;
In the discussion of Win32 mutexes, we saw that calling WaitFor-
SingleObject(hMutex, 0) would test hMutex right away and return. By
examining the return value and comparing it to WAIT_TIMEOUT, we can tell
whether the mutex was locked. The Pthreads library has a similar function,
pthread_mutex_trylock(&mutex), which tests the mutex to see whether
its locked and then returns. If it returns EBUSY, the mutex is already locked.
Its important to note in both the Windows and Pthreads version of this
function, if the mutex is unlocked, this call will lock it. It behooves you
therefore to check the return value, so as to avoid inadvertently locking a
mutex simply because you were trying to see whether it was available. It is
expected that you will use this test-and-lock behavior in situations where
you would like to lock a mutex, but if the mutex is already locked, you
might want to perform other activities before testing the mutex again.

Signaling
Many multi-threading programmers find the event model of
communication error prone. As a result, certain APIs exclude them. The
Pthreads model has no direct counterpart to the Windows concept of
 Chapter 5: Threading APIs 125

events. Rather, two separate constructs can be used to achieve the same
ends. They are condition variables and the semaphore.

Condition Variables
A condition variable is a mechanism that is tightly bound to a mutex and
a data item. It is used when one or more threads are waiting for the value
of the data item to change. Rather than spinning, the threads block on
the condition variable and wait for it to be signaled by some other thread.
This signal notifies the waiting threads that the data item has changed
and enables the threads to begin or resume processing.
This works in a very mechanical way. The data item is declared, for
instance, with a flag that tells a consumer thread that the producer thread
has data ready for it, and that the data is protected by a mutex. The data
item and the mutex together are associated with a condition variable.
When the producer thread changes the flag, after unlocking and
relocking the mutex, it signals the condition variable, which announces
that the flag has changed value. This announcement can be sent
optionally to a single thread or broadcast to all threads blocking on the
condition variable. In addition to the announcement, the signal unblocks
the waiting thread or threads.
Listing 5.13 illustrates how this works by showing two threads
waiting on a condition variable. The listing is somewhat longer than the
others presented in this book, but it shows how to address a very typical
problem in programming with threads.

1 #include <stdio.h>
2 #include <stdlib.h>
3
4 #include <pthread.h>
5
6 #define BLOCK_SIZE 100
7 #define BUF_SIZE 1000000
8
9 size_t bytesRead;
10
11 typedef struct {
12 pthread_mutex_t mutex; // mutex
13 pthread_cond_t cv; // condition variable
14 int data; // data item used as a flag
15 } flag;
16
17 flag ourFlag = { // default initialization
18 PTHREAD_MUTEX_INITIALIZER,
126 Multi-Core Programming

19 PTHREAD_COND_INITIALIZER,
20 0 }; // data item set to 0
21
22 pthread_t hThread1, hThread2; // the waiting threads
23 void* PrintCountRead( void* ); // the thread function
24
25 int main( int argc, char *argv[] )
26 {
27 FILE *infile;
28 char *inbuf;
29 int status;
30
31 if ( argc != 2 )
32 {
33 printf( "Usage GetSetEvents filename\n" );
34 return( -1 );
35 }
36
37 infile = fopen( argv[1], "r+b" );
38 if ( infile == NULL )
39 {
40 printf( "Error opening %s\n", argv[1] );
41 return( -1 );
42 }
43
44 inbuf = (char*) malloc ( BUF_SIZE );
45 if ( inbuf == NULL )
46 {
47 printf( "Could not allocate read buffer\n" );
48 return( -1 );
49 }
50
51 // now start up two threads
52 pthread_create( &hThread1, NULL,
53 PrintCountRead, (void *) NULL );
54 pthread_create( &hThread2, NULL,
55 PrintCountRead, (void *) NULL );
56
57 bytesRead = fread( inbuf, 1, BLOCK_SIZE, infile );
58 if ( bytesRead < BLOCK_SIZE )
59 {
60 printf( "Need a file longer than %d bytes\n",
61 BLOCK_SIZE );
62 return( -1 );
63 }
64 else // now we tell the waiting thread(s)
65 {
66 // first, lock the mutex
67 status = pthread_mutex_lock( &ourFlag.mutex );
68 if ( status != 0 )
 Chapter 5: Threading APIs 127

69 {
70 printf( "error locking mutex in main func.\n" );
71 exit( -1 );
72 }
73
74 ourFlag.data = 1; // change the data item
75 // then broadcast the change
76 status = pthread_cond_broadcast( &ourFlag.cv ) ;
77 if ( status != 0 )
78 {
79 printf( "error broadcasting condition var\n" );
80 exit( -1 );
81 }
82
83 // unlock the mutex
84 status = pthread_mutex_unlock( &ourFlag.mutex );
85 if ( status != 0 )
86 {
87 printf( "error unlocking mutex in waiting \
88 function\n" );
89 exit( -1 );
90 }
91 }
92
93 while ( !feof( infile ) &&
94 bytesRead < BUF_SIZE - BLOCK_SIZE )
95 bytesRead += fread(inbuf, 1, BLOCK_SIZE, infile );
96
97 printf("Read a total of %d bytes\n", (int)bytesRead);
98 return( 0 );
99 }
100
101 // the thread function, which waits on the
102 // condition variable
103 void *PrintCountRead( void* pv )
104 {
105 int status;
106
107 // lock the mutex
108 status = pthread_mutex_lock( &ourFlag.mutex );
109 if ( status != 0 )
110 {
111 printf( "error locking mutex in waiting func.\n" );
112 exit( -1 );
113 }
114
115 // now wait on the condition variable
116 // (loop should spin once only)
117 while ( ourFlag.data == 0 )
118 {
128 Multi-Core Programming

119 status = pthread_cond_wait( &ourFlag.cv,

120 &ourFlag.mutex );
121 if ( status != 0 )
122 {
123 printf("error waiting on condition variable\n");
124 exit( -1 );
125 }
126 }
127
128 if ( ourFlag.data != 0 )
129 {
130 printf( "Condition was signaled. "
131 "Main thread has read %06d bytes\n",
132 (int) bytesRead );
133 }
134
135 // unlock the mutex
136 status = pthread_mutex_unlock( &ourFlag.mutex );
137 if ( status != 0 )
138 {
139 printf("error unlocking mutex in waiting func.\n");
140 exit( -1 );
141 }
142
143 return( pv );
144 }

Listing 5.13 Waking Two Threads through a Broadcast to Condition Variables in

Pthreads

Several subtleties come into play. Before examining them, lets go

over what this code does. At a global level (lines 1121), the code creates
a structure that includes a mutex, a condition variable, and a data item
that serves as a flag. This code also initializes both the mutex and the
condition variable to the Pthreads defaults, and 0 for the flag.
The code opens and reads a file that is specified on the command
line. Then, two threads are created (lines 5155); they both use the same
thread function, PrintCountRead(). This function locks the mutex
(lines 108113) and then starts a loop that waits on the value of the flag.
The function then calls pthread_cond_wait() (lines 119120), which
is the code that registers with the system that it wants to be awakened
when the condition variable is signaled. At this point, the thread blocks
while waiting for a signal. When the signal arrives, the thread wakes up
and proceeds. The condition variable is signaled from the main function
after the flags value is changed to 1 (lines 7476). The loop condition
 Chapter 5: Threading APIs 129

now proves false, and execution flows to the next statement. Here, the
flags value is checked again (line 128130) and the dependent action
printing the number of bytes read by the principal threadis performed.
The mutex is then unlocked (lines 135141) and the worker thread exits.
After starting up the two worker threads, which are both blocked
waiting for their condition variables to be signaled, the main thread reads
one buffer of data (line 57). When this read is successful, it signals the
worker threads that they can proceed. It does this by locking the mutex
and broadcasting the signal to all waiting threads via
pthread_cond_broadcast() (line 76). It then unlocks the mutex and
finishes reading the file. This routine could have instead used
pthread_cond_signal() to emit the signal. However, that call would
have signaled only one waiting thread, rather than all of them. Such an
option would be useful if several waiting threads are all waiting to do the
same thing, but the desired activity cannot be parallelized.
The program in Listing 5.14 generates the following output when run
on a file consisting of 435,676 bytes.
Condition was signaled. Main thread has read 002700 bytes
Condition was signaled. Main thread has read 011200 bytes
Read a total of 435676 bytes
You might be tempted to use condition variables without the
required mutex. This will lead to problems. Pthreads is designed to use a
mutex with condition variables, as can be seen in the parameters in
pthread_cond_wait(), which takes a pointer to the condition variable
and one to the mutex. In fact, without the mutex, the code will not
compile properly. The mutex is needed by the Pthreads architecture to
correctly record the occurrence of the signal used by the condition
variable.
The code in Listing 5.14 is typical of producer/consumer situations.
In those, typically, the program starts up a bunch of threads. The
producer threadsin this case, the one reading the filemust generate
data or actions for the consumer or worker threads to process. Typically,
the consumer threads are all suspended pending a signal sent when there
is data to consume. In this situation, .NET implements handles via a
thread pool; however, Pthreads has no built-in thread pool mechanism.

Semaphores
The semaphore is comparable to those in the Win32 APIs, described
earlier. A semaphore is a counter that can have any nonnegative value.
Threads wait on a semaphore. When the semaphores value is 0, all
130 Multi-Core Programming

threads are forced to wait. When the value is nonzero, a waiting thread is
released to work. The thread that gets released is determined first by
thread priority, then by whoever attached to the semaphore first. When
a thread releases, that is, becomes unblocked, it decrements the value of
the semaphore. In typical constructs, the semaphore is set to 0
(blocking), which forces dependent threads to wait. Another thread
increments the semaphore; this process is known as posting. One
waiting thread is thereby released and in releasing, it decrements the
semaphore back to 0. This blocks all other threads still waiting on
the semaphore. This design makes the semaphore a convenient way to
tell a single waiting thread that it has work to be performed.
Technically speaking, Pthreads does not implement semaphores; they
are a part of a different POSIX specification. However, semaphores are
used in conjunction with Pthreads thread-management functionality, as
you shall see presently. Listing 5.14 illustrates the use of Pthreads with
semaphores. The program reads a file and signals another thread to print
the count of bytes read. Nonessential parts of the listing have been
removed.

1 #include <stdio.h>
2 #include <stdlib.h>
3
4 #include <pthread.h>
5 #include <semaphore.h>
6
7 #define BLOCK_SIZE 100
8 #define BUF_SIZE 1000000
9
10 size_t bytesRead;
11
12 sem_t sReadOccurred; // the semaphore we'll use
13 pthread_t hThread; // the waiting thread
14 void*
15 PrintCountRead( void* ); // the thread function
16
17 int main( int argc, char *argv[] )
18 {
19 . . . open the input file here. . .
20
21 // first initialize the semaphore
22 sem_init( &sReadOccurred, // address of the semaphore
23 0, // 0 = share only with threads in this program
24 0 ); // initial value. 0 = make threads wait
25
26 // now start up the thread
 Chapter 5: Threading APIs 131

27 pthread_create(
28 &hThread,
29 NULL,
30 PrintCountRead,
31 (void *) NULL );
32
33 bytesRead = fread( inbuf, 1, BLOCK_SIZE, infile );
34 if ( bytesRead < BLOCK_SIZE )
35 {
36 printf( "Need a file longer than %d bytes\n",
37 BLOCK_SIZE );
38 return( -1 );
39 }
40 else
41 sem_post( &sReadOccurred ); // release the
42 // waiting threads
43
44 . . . finish reading file and print total bytes read. . .
45
46 return( 0 );
47 }
48
49 // the thread function, which waits for the event before
50 // proceeding
51 void *PrintCountRead( void* pv )
52 {
53 int i;
54
55 sem_wait( &sReadOccurred ); // wait on the semaphore
56 printf( "Have now read %06d bytes\n",
57 (int) bytesRead );
58 return( pv );
59 }

Listing 5.14 Using a Pthreads Semaphore to Indicate Program Status

The code opens an input file, creates a semaphore (lines 2125)

and starts up a thread (lines 2632) before it reads any data. Notice that
the semaphore refers to the semaphore data item on line 12. In line 55,
we see that the created thread, which will report the number of bytes
read, is waiting for the semaphore to signal. After the first read occurs,
the semaphore is signaled (line 41) and releases the waiting thread to
begin reporting the number of input bytes read. Even though the
reporting thread was started up prior to any reads, it is incapable of
reporting 0 bytes read because it has blocked on the semaphore until
after the first read completed.
132 Multi-Core Programming

This use of a semaphore is valuable in the producer-consumer model.

The consumer thread function is set up to contain an infinite loop, such
as with while(1). When the producer has placed data in a data
structure, generally a queue, it signals the semaphore, thereby releasing
the consumer threads main loop. This release resets the semaphore;
once the consumer routine is complete and loops anew, it will block on
the semaphore until the producer thread releases it again. If the producer
thread should post to the semaphore while the consumer thread is
working, the thread will discover upon looping that the semaphore is
already unlocked and so it will continue processing without stopping. By
this means, the producer thread can directly control when and how often
the consumer thread performs its work.
The difficulty with semaphores is that they are limited when dealing
with multiple consumer threads. To solve this problem, developers using
the Pthreads library rely on condition variables.

Compilation and Linking

Pthreads code should include the pthread.h header file. On the
compilation command line, the Pthreads library should be specified to
the linker on UNIX and Linux environments using the -lpthread
command-line switch. For the pthreads-win32 version mentioned
earlier, include the bundled header file and link to the DLL, and the
Pthreads programs will compile and run correctly on Windows.

Key Points
This chapter provided an overview of two threading APIs: the Microsoft
Windows model, and the POSIX threads (Pthreads) model. When
developing applications based on these APIs, you should keep the
following points in mind:
Multi-threaded applications targeting Microsoft Windows can be
written in either native or managed code.
Since the CreateThread() function does not perform per-thread
initialization of C runtime datablocks and variables, you cannot
reliably use CreateThread() in any application that uses the C
runtime library. Use _beginthreadex() function instead.
Thread termination should be handled very carefully. Avoid using
functions such as TerminateThread().
 Chapter 5: Threading APIs 133

Threads can communicate with one another using Events.

Thread synchronization may be accomplished through the use of
Mutexes, Semaphores, CriticalSections, and Interlocked functions.
Thread pool support is built into the API.
Windows supports multiple thread-priority levels.
Processor affinity is a mechanism that allows the programmer to
specify which processor a thread should try to run on.
POSIX threads (Pthreads) is a portable threading API that is
supported on a number of platforms.
Different platforms support different Pthreads capabilities.
Features may not be available in all Pthreads environments.
 Chapter 6

OpenMP :
A Portable Solution
for Threading
T he major CPU vendors are shifting gears, choosing to add parallelism
support on-chip with multi-core processors in order to avoid many of
the technological hurdles in boosting speeds, while still offering a better
performing processor. However, if your software does not take
advantage of these multiple cores, it may not run any faster. That is

where OpenMP plays a key role by providing an easy method for
threading applications without burdening the programmer with the
complications of creating, synchronizing, load balancing, and destroying
threads.
The OpenMP standard was formulated in 1997 as an API for writing
portable, multithreaded applications. It started as a Fortran-based
standard, but later grew to include C and C++. The current version is
OpenMP Version 2.5, which supports Fortran, C, and C++. Intel C++
and Fortran compilers support the OpenMP Version 2.5 standard
(www.openmp.org). The OpenMP programming model provides a
platform-independent set of compiler pragmas, directives, function calls,
and environment variables that explicitly instruct the compiler how and
where to use parallelism in the application. Many loops can be threaded
by inserting only one pragma right before the loop, as demonstrated by
examples in this chapter. By leaving the nitty-gritty details to the
compiler and OpenMP runtime library, you can spend more time
determining which loops should be threaded and how to best restructure
the algorithms for performance on multi-core processors. The full
potential of OpenMP is realized when it is used to thread the most time-
consuming loops, that is, the hot spots.

135
136 Multi-Core Programming

Tackling the topic of OpenMP in a single chapter is an intimidating

task. Therefore, this chapter serves as a bridge for you, allowing you to
reach a point where you have a fundamental understanding of threading
with OpenMP from which you can build your broader practical
knowledge. The power and simplicity of OpenMP can be demonstrated
by looking at an example. The following loop converts each 32-bit RGB
(red, green, blue) pixel in an array into an 8-bit grayscale pixel. The one
pragma, which has been inserted immediately before the loop, is all that
is needed for parallel execution under OpenMP.
#pragma omp parallel for
for ( i = 0; i < numPixels; i++)
{
pGrayScaleBitmap[i] = (unsigned BYTE)
( pRGBBitmap[i].red * 0.299 +
pRGBBitmap[i].green * 0.587 +
pRGBBitmap[i].blue * 0.114 );
}

Lets take a closer look at the loop. First, the example uses work-sharing,
which is the general term that OpenMP uses to describe distributing
work across threads. When work-sharing is used with the for construct,
as shown in this example, the iterations of the loop are distributed
among multiple threads. The OpenMP implementation determines how
many threads to create and how best to manage them. All the
programmer needs to do is to tell OpenMP which loop should be
threaded. No need for programmers to add a lot of codes for creating,
initializing, managing, and killing threads in order to exploit parallelism.
OpenMP compiler and runtime library take care of these and many other
details behind the scenes.
In the current OpenMP specification Version 2.5, OpenMP places the
following five restrictions on which loops can be threaded:
The loop variable must be of type signed integer. Unsigned
integers will not work. Note: this restriction is to be removed in
the future OpenMP specification Version 3.0.
The comparison operation must be in the form loop_variable
<, <=, >, or >= loop_invariant_integer.
The third expression or increment portion of the for loop must
be either integer addition or integer subtraction and by a loop-
invariant value.
 Chapter 6: OpenMP: A Portable Solution for Threading 137

If the comparison operation is < or <=, the loop variable must

increment on every iteration; conversely, if the comparison
operation is > or >=, the loop variable must decrement on every
iteration.
The loop must be a single entry and single exit loop, meaning no
jumps from the inside of the loop to the outside or outside to the
inside are permitted with the exception of the exit statement,
which terminates the whole application. If the statements goto
or break are used, they must jump within the loop, not outside
it. The same goes for exception handling; exceptions must be
caught within the loop.
Although these restrictions may sound somewhat limiting, most loops
can easily be rewritten to conform to them. The restrictions listed above
must be observed so that the compiler can parallelize loops via OpenMP.
However, even when the compiler parallelizes the loop, you must still
ensure the loop is functionally correct by watching out for the issues in
the next section.

Challenges in Threading a Loop

Threading a loop is to convert independent loop iterations to threads and
run these threads in parallel. In some sense, this is a re-ordering
transformation in which the original order of loop iterations can be
converted to into an undetermined order. In addition, because the loop
body is not an atomic operation, statements in the two different
iterations may run simultaneously. In theory, it is valid to convert a
sequential loop to a threaded loop if the loop carries no dependence.
Therefore, the first challenge for you is to identify or restructure the hot
loop to make sure that it has no loop-carried dependence before adding
OpenMP pragmas.

Loop-carried Dependence
Even if the loop meets all five loop criteria and the compiler threaded
the loop, it may still not work correctly, given the existence of data
dependencies that the compiler ignores due to the presence of OpenMP
pragmas. The theory of data dependence imposes two requirements
138 Multi-Core Programming

that must be met for a statement S2 and to be data dependent on

statement S1.
There must exist a possible execution path such that statement S1
and S2 both reference the same memory location L.
The execution of S1 that references L occurs before the execution
of S2 that references L.
In order for S2 to depend upon S1, it is necessary for some execution of S1
to write to a memory location L that is later read by an execution of S2.
This is also called flow dependence. Other dependencies exist when two
statements write the same memory location L, called an output
dependence, or a read occurs before a write, called an anti-dependence.
This pattern can occur in one of two ways:
S1 can reference the memory location L on one iteration of a loop;
on a subsequent iteration S2 can reference the same memory
location L.
S1 and S2 can reference the same memory location L on the same
loop iteration, but with S1 preceding S2 during execution of the
loop iteration.
The first case is an example of loop-carried dependence, since the
dependence exists when the loop is iterated. The second case is an
example of loop-independent dependence; the dependence exists
because of the position of the code within the loops. Table 6.1 shows
three cases of loop-carried dependences with dependence distance d,
where 1 d n, and n is the loop upper bound.
Table 6.1 The Different Cases of Loop-carried Dependences
iteration k iteration k + d

Loop-carried flow dependence

statement S1 write L
statement S2 read L
Loop-carried anti-dependence
statement S1 read L
statement S2 write L
Loop-carried output dependence
statement S1 write L
statement S2 write L
 Chapter 6: OpenMP: A Portable Solution for Threading 139

Lets take a look at the following example where d = 1 and n = 99.

The write operation is to location x[k] at iteration k in S1, and a read
from it at iteration k+1 in S2, thus a loop-carried flow dependence occurs.
Furthermore, with the read from location y[k1] at iteration k in S1, a
write to it is performed at iteration k+1 in S2, hence, the loop-carried anti-
dependence exists. In this case, if a parallel for pragma is inserted for
threading this loop, you will get a wrong result.
// Do NOT do this. It will fail due to loop-carried
// dependencies.

x[0] = 0;
y[0] = 1;
#pragma omp parallel for private(k)

for ( k = 1; k < 100; k++ ) {

x[k] = y[k-1] + 1; // S1
y[k] = x[k-1] + 2; // S2
}
Because OpenMP directives are commands to the compiler, the compiler
will thread this loop. However, the threaded code will fail because of
loop-carried dependence. The only way to fix this kind of problem is to
rewrite the loop or to pick a different algorithm that does not contain the
loop-carried dependence. With this example, you can first predetermine
the initial value of x[49] and y[49]; then, you can apply the loop strip-
mining technique to create a loop-carried dependence-free loop m.
Finally, you can insert the parallel for to parallelize the loop m. By
applying this transformation, the original loop can be executed by two
threads on a dual-core processor system.
// Effective threading of the loop using strip-mining
// transformation.

x[0] = 0;
y[0] = 1;
x[49] = 74; //derived from the equation x(k)=x(k-2)+3
y[49] = 74; //derived from the equation y(k)=y(k-2)+3

#pragma omp parallel for private(m, k)

for (m=0, m<2; m++) {

for ( k = m*49+1; k < m*50+50; k++ ) {
x[k] = y[k-1] + 1; // S1
y[k] = x[k-1] + 2; // S2
}
}
140 Multi-Core Programming

Besides using the parallel for pragma, for the same example, you
can also use the parallel sections pragma to parallelize the original
loop that has loop-carried dependence for a dual-core processor system.
// Effective threading of a loop using parallel sections

#pragma omp parallel sections private(k)

{ { x[0] = 0; y[0] = 1;
for ( k = 1; k < 49; k++ ) {
x[k] = y[k-1] + 1; // S1
y[k] = x[k-1] + 2; // S2
}
}

#pragma omp section

{ x[49] = 74; y[49] = 74;
for ( k = 50; k < 100; k++ ) {
x[k] = y[k-1] + 1; // S3
y[k] = x[k-1] + 2; // S4
}
}
}

With this simple example, you can learn several effective methods
from the process of parallelizing a loop with loop-carried dependences.
Sometimes, a simple code restructure or transformation is necessary to
get your code threaded for taking advantage of dual-core and multi-core
processors besides simply adding OpenMP pragmas.

Data-race Conditions
Data-race conditions that are mentioned in the previous chapters could
be due to output dependences, in which multiple threads attempt
to update the same memory location, or variable, after threading. In
general, the OpenMP C++ and Fortran compilers do honor OpenMP
pragmas or directives while encountering them during compilation
phase, however, the compiler does not perform or ignores the detection
of data-race conditions. Thus, a loop similar to the following example, in
which multiple threads are updating the variable x will lead to
undesirable results. In such a situation, the code needs to be modified
via privatization or synchronized using mechanisms like mutexes. For
example, you can simply add the private(x) clause to the parallel
for pragma to eliminate the data-race condition on variable x for
this loop.
 Chapter 6: OpenMP: A Portable Solution for Threading 141

// A data race condition exists for variable x;

// you can eliminate it by adding private(x) clause.

#pragma omp parallel for

for ( k = 0; k < 80; k++ )
{
x = sin(k*2.0)*100 + 1;
if ( x > 60 ) x = x % 60 + 1;
printf ( "x %d = %d\n", k, x );
}
As discussed previously, data-race conditions can sometimes be
difficult to spot; that is, more difficult than shown in this example. When
using the full thread synchronization tools in the Windows API or in
Pthreads, developers are more likely to avoid these issues because data is
designed from the start to be managed with threading contention and
race conditions in mind. However, when using OpenMP, it is easier to
overlook data-race conditions. One tool that helps identify such situations

is Intel Thread Checker, which is an add-on to Intel VTune
Performance Analyzer. Intel Thread Checker is discussed in more detail
in Chapter 11.

Managing Shared and Private Data

In writing multithreaded programs, understanding which data is shared
and which is private becomes extremely important, not only to
performance, but also for program correctness. OpenMP makes this
distinction apparent to the programmer through a set of clauses such as
shared, private, and default, and it is something that you can set
manually. With OpenMP, it is the developers responsibility to indicate to
the compiler which pieces of memory should be shared among the
threads and which pieces should be kept private. When memory is
identified as shared, all threads access the exact same memory location.
When memory is identified as private, however, a separate copy of the
variable is made for each thread to access in private. When the loop
exits, these private copies become undefined.
By default, all the variables in a parallel region are shared, with three
exceptions. First, in parallel for loops, the loop index is private. In
the next example, the k variable is private. Second, variables that are
local to the block of the parallel region are private. And third, any
variables listed in the private, firstprivate, lastprivate, or
reduction clauses are private. The privatization is done by making a
distinct copy of each of these variables for each thread.
142 Multi-Core Programming

Each of the four clauses takes a list of variables, but their semantics are
all different. The private clause says that each variable in the list should
have a private copy made for each thread. This private copy is initialized
with its default value, using its default constructor where appropriate. For
example, the default value for variables of type int is 0. In OpenMP,
memory can be declared as private in the following three ways.
Use the private, firstprivate, lastprivate, or reduction
clause to specify variables that need to be private for each
thread.
Use the threadprivate pragma to specify the global variables
that need to be private for each thread.
Declare the variable inside the loopreally inside the OpenMP
parallel regionwithout the static keyword. Because static
variables are statically allocated in a designated memory area by
the compiler and linker, they are not truly private like other
variables declared within a function, which are allocated within
the stack frame for the function.
The following loop fails to function correctly because the variable x is
shared. It needs to be private. Given example below, it fails due to the
loop-carried output dependence on the variable x. The x is shared among
all threads based on OpenMP default shared rule, so there is a data-race
condition on the x while one thread is reading x, another thread might
be writing to it
#pragma omp parallel for
for ( k = 0; k < 100; k++ ) {
x = array[k];
array[k] = do_work(x);
}

This problem can be fixed in either of the following two ways, which
both declare the variable x as private memory.
// This works. The variable x is specified as private.

#pragma omp parallel for private(x)

for ( k = 0; k < 100; k++ )
{
x = array[i];
array[k] = do_work(x);
}
 Chapter 6: OpenMP: A Portable Solution for Threading 143

// This also works. The variable x is now private.

#pragma omp parallel for
for ( k = 0; k < 100; k++ )
{
int x; // variables declared within a parallel
// construct are, by definition, private
x = array[k];
array[k] = do_work(x);
}

Every time you use OpenMP to parallelize a loop, you should carefully
examine all memory references, including the references made by called
functions.

Loop Scheduling and Partitioning

To have good load balancing and thereby achieve optimal performance
in a multithreaded application, you must have effective loop scheduling
and partitioning. The ultimate goal is to ensure that the execution cores
are busy most, if not all, of the time, with minimum overhead of
scheduling, context switching and synchronization. With a poorly
balanced workload, some threads may finish significantly before
others, leaving processor resources idle and wasting performance
opportunities. In order to provide an easy way for you to adjust the
workload among cores, OpenMP offers four scheduling schemes that
are appropriate for many situations: static, dynamic, runtime, and
guided. The Intel C++ and Fortran compilers support all four of these
scheduling schemes.
A poorly balanced workload is often caused by variations in compute
time among loop iterations. It is usually not too hard to determine the
variability of loop iteration compute time by examining the source code.
In most cases, you will see that loop iterations consume a uniform
amount of time. When thats not true, it may be possible to find a set of
iterations that do consume similar amounts of time. For example,
sometimes the set of all even iterations consumes about as much time as
the set of all odd iterations, or the set of the first half of the loop
consumes about as much time as the second half. On the other hand, it
may be impossible to find sets of loop iterations that have a uniform
execution time. In any case, you can provide loop scheduling
information via the schedule(kind [, chunksize]) clause, so that the
compiler and runtime library can better partition and distribute the
iterations of the loop across the threads, and therefore the cores, for
optimal load balancing.
144 Multi-Core Programming

By default, an OpenMP parallel for or worksharing for loop uses

static-even scheduling. This means the iterations of a loop are distributed
among the threads in a roughly equal number of iterations. If m iterations
and N threads are in the thread team, each thread gets m/N iterations,
and the compiler and runtime library correctly handles the case when m
is not evenly divisible by N.
With the static-even scheduling scheme, you could minimize the
chances of memory conflicts that can arise when more than one
processor is trying to access the same piece of memory. This approach
is workable because loops generally touch memory sequentially, so
splitting up the loop into large chunks results in little chance of
overlapping memory and a reasonable chance of good processor cache
efficiency. Consider the following simple loop when executed using
static-even scheduling and two threads.
#pragma omp parallel for
for ( k = 0; k < 1000; k++ ) do_work(k);
OpenMP will execute loop iterations 0 to 499 on one thread and
500 to 999 on the other thread. While this partition of work might be a
good choice for memory issues, it could be bad for load balancing.
Unfortunately, the converse is also true: what might be good for load
balancing could be bad for memory performance. Therefore,
performance engineers must strike a balance between optimal memory
usage and optimal load balancing by measuring performance to see what
method produces the best results.
Loop-scheduling and partitioning information is conveyed to the
compiler and runtime library on the OpenMP for construct with the
schedule clause.
#pragma omp for schedule(kind [, chunk-size])

The four schedule schemes specified in the OpenMP standard are

summarized in Table 6.2. The optional parameter chunk-size, when
specified, must be a loop-invariant positive integer constant or integer
expression.
Be careful when you adjust the chunk size, because performance
can be adversely affected. As the chunk size shrinks, the number of
times a thread needs to retrieve work from the work queue increases.
As a result, the overhead of going to the work queue increases, thereby
reducing performance and possibly offsetting the benefits of load
balancing.
 Chapter 6: OpenMP: A Portable Solution for Threading 145

Table 6.2 The Four Schedule Schemes in OpenMP

Schedule Type Description
static (default Partitions the loop iterations into equal-sized chunks or as
with no chunk nearly equal as possible in the case where the number of loop
size) iterations is not evenly divisible by the number of threads
multiplied by the chunk size. When chunk size is not specified,
the iterations are divided as evenly as possible, with one chunk
per thread. Set chunk to 1 to interleave the iterations.
dynamic Uses an internal work queue to give a chunk-sized block of loop
iterations to each thread as it becomes available. When a thread
is finished with its current block, it retrieves the next block of
loop iterations from the top of the work queue. By default, chunk
size is 1. Be careful when using this scheduling type because of
the extra overhead required.
guided Similar to dynamic scheduling, but the chunk size starts off large
and shrinks in an effort to reduce the amount of time threads
have to go to the work queue to get more work. The optional
chunk parameter specifies the minimum size chunk to use,
which, by default, is 1.
runtime Uses the OMP_SCHEDULE environment variable at runtime to
specify which one of the three loop-scheduling types should be
used. OMP_SCHEDULE is a string formatted exactly the same
as it would appear on the parallel construct.

For dynamic scheduling, the chunks are handled with the first-come,
first-serve scheme, and the default chunk size is 1. Each time, the number
of iterations grabbed is equal to the chunk size specified in the schedule
clause for each thread, except the last chunk. After a thread has finished
executing the iterations given to it, it requests another set of chunk-size
iterations. This continues until all of the iterations are completed. The
last set of iterations may be less than the chunk size. For example, if
the chunk size is specified as 16 with the schedule(dynamic,16)
clause and the total number of iterations is 100, the partition would be
16,16,16,16,16,16,4 with a total of seven chunks.
For guided scheduling, the partitioning of a loop is done based on the
following formula with a start value of 0 = number of loop iterations.

k = k
2N
where N is the number of threads, k denotes the size of the kth chunk,
starting from the 0th chunk, and k denotes the number of remaining
unscheduled loop iterations while computing the size of kth chunk.
146 Multi-Core Programming

When k gets too small, the value gets clipped to the chunk size S
that is specified in the schedule (guided, chunk-size) clause. The
default chunk size setting is 1, if it is not specified in the schedule
clause. Hence, for the guided scheduling, the way a loop is partitioned
depends on the number of threads (N), the number of iterations (0) and
the chunk size (S).
For example, given a loop with 0 = 800, N = 2, and S = 80, the loop
partition is {200, 150, 113, 85, 80, 80, 80, 12}. When 4 is smaller than
80, it gets clipped to 80. When the number of remaining unscheduled
iterations is smaller than S, the upper bound of the last chunk is trimmed
whenever it is necessary. The guided scheduling supported in the Intel
C++ and Fortran compilers are a compliant implementation specified in
the OpenMP Version 2.5 standard.
With dynamic and guided scheduling mechanisms, you can tune your
application to deal with those situations where each iteration has variable
amounts of work or where some cores (or processors) are faster than
others. Typically, guided scheduling performs better than dynamic
scheduling due to less overhead associated with scheduling.
The runtime scheduling scheme is actually not a scheduling scheme
per se. When runtime is specified in the schedule clause, the OpenMP
runtime uses the scheduling scheme specified in the OMP_SCHEDULE
environment variable for this particular for loop. The format for the
OMP_SCHEDULE environment variable is schedule-type[,chunk-size].
For example:
export OMP_SCHEDULE=dynamic,16

Using runtime scheduling gives the end-user some flexibility in

selecting the type of scheduling dynamically among three previously
mentioned scheduling mechanisms through the OMP_SCHEDULE
environment variable, which is set to static by default.
Furthermore, understanding the loop scheduling and partitioning
schemes will significantly help you to choose the right scheduling
scheme, help you to avoid false-sharing for your applications at runtime,
and lead to good load balancing. Considering the following example:
float x[1000], y[1000];
#pragma omp parallel for schedule(dynamic, 8)
for (k=0; k<1000; k++) {
x[k] = cos(k)* x[k] + sin(k) * y[k]
}
 Chapter 6: OpenMP: A Portable Solution for Threading 147

Assume you have a dual-core processor system and the cache line size
is 64 bytes. For the sample code shown above, two chunks (or array
sections) can be in the same cache line because the chunk size is set to 8
in the schedule clause. So each chunk of array x takes 32 bytes per cache
line, which leads to two chunks placed in the same cache line. Because
two chunks can be read and written by two threads at the same time, this
will result in many cache line invalidations, although two threads do not
read/write the same chunk. This is called false-sharing, as it is not
necessary to actually share the same cache line between two threads. A
simple tuning method is to use schedule(dynamic,16), so one chunk
takes the entire cache line to eliminate the false-sharing. Eliminating false-
sharing through the use of a chunk size setting that is aware of cache line
size will significantly improve your application performance.

Effective Use of Reductions

In large applications, you can often see the reduction operation inside
hot loops. Loops that reduce a collection of values to a single value are
fairly common. Consider the following simple loop that calculates the
sum of the return value of the integer-type function call func(k) with
the loop index value as input data.
sum = 0;
for ( k = 0; k < 100; k++ ){
sum = sum + func(k); // func has no side-effects
}

It looks as though the loop-carried dependence on sum would prevent

threading. However, if you have a dual-core processor system, you can
perform the privatizationthat is, create a stack variable temp from
which memory is allocated from automatic storage for each threadand
perform loop partitioning to sum up the value of two sets of calls in
parallel, as shown in the following example.

Thread 0: Thread 1:

temp = 0; temp = 0;
for (k=0; k<50; k++) { for (k=50; k<100; k++) {
temp = temp + func(k); temp = temp + func(k)
} }

lock (&sum) lock(&sum)

sum = sum + temp sum = sum + temp
unlock (&sum) unlock(&sum)
148 Multi-Core Programming

At the synchronization point, you can combine the partial sum results
from each thread to generate the final sum result. In order to perform
this form of recurrence calculation in parallel, the operation must be
mathematically associative and commutative. You may notice that the
variable sum in the original sequential loop must be shared to guarantee
the correctness of the multithreaded execution of the loop, but it also
must be private to permit access by multiple threads using a lock or a
critical section for the atomic update on the variable sum to avoid data-
race condition. To solve the problem of both sharing and protecting sum
without using a lock inside the threaded loop, OpenMP provides the
reduction clause that is used to efficiently combine certain associative
arithmetical reductions of one or more variables in a loop. The following
loop uses the reduction clause to generate the correct results.
sum = 0;
#pragma omp parallel for reduction(+:sum)
for (k = 0; k < 100; k++) {
sum = sum + func(k);
}

Given the reduction clause, the compiler creates private copies of the
variable sum for each thread, and when the loop completes, it adds the
values together and places the result in the original variable sum.
Other reduction operators besides + exist. Table 6.3 lists those C++
reduction operators specified in the OpenMP standard, along with the initial
valueswhich are also the mathematical identity valuefor the temporary
private variables. You can also find a list of Fortran reduction operators along
with their initial values in OpenMP specification Version 2.5.

Table 6.3 Reduction Operators and Reduction Variables Initial Value in

OpenMP
Operator Initialization Value
+ (addition) 0
- (subtraction) 0
* (multiplication) 1
& (bitwise and) ~0
| (bitwise or) 0
^ (bitwise exclusive or) 0
&& (conditional and) 1
|| (conditional or) 0
 Chapter 6: OpenMP: A Portable Solution for Threading 149

For each variable specified in a reduction clause, a private copy is

created, one for each thread, as if the private clause is used. The
private copy is then initialized to the initialization value for the operator,
as specified in Table 6.3. At the end of the region or the loop for which
the reduction clause was specified, the original reduction variable is
updated by combining its original value with the final value of each of
the private copies, using the operator specified. While identifying the
opportunities to explore the use of the reduction clause for threading,
you should keep the following three points in mind.
The value of the original reduction variable becomes undefined
when the first thread reaches the region or loop that specifies the
reduction clause and remains so until the reduction
computation is completed.
If the reduction clause is used on a loop to which the nowait is
also applied, the value of original reduction variable remains
undefined until a barrier synchronization is performed to ensure
that all threads have completed the reduction.
The order in which the values are combined is unspecified.
Therefore, comparing sequential and parallel runs, even between
two parallel runs, does not guarantee that bit-identical results will
be obtained or that side effects, such as floating-point exceptions,
will be identical.

Minimizing Threading Overhead

Using OpenMP, you can parallelize loops, regions, and sections or
straight-line code blocks, whenever dependences do not forbids them
being executed in parallel. In addition, because OpenMP employs the
simple fork-join execution model, it allows the compiler and run-time
library to compile and run OpenMP programs efficiently with lower
threading overhead. However, you can improve your application
performance by further reducing threading overhead.
Table 6.4 provides measured costs of a set of OpenMP constructs and

clauses on a 4-way Intel Xeon processor-based system running at
3.0 gigahertz with the Intel compiler and runtime library. You can see that
the cost for each construct or clause is small. Most of them are less than
7 microseconds except the schedule(dynamic) clause. The schedule
(dynamic) clause takes 50 microseconds, because its default chunk size is
1, which is too small. If you use schedule(dynamic,16), its cost is
150 Multi-Core Programming

reduced to 5.0 microseconds. Note that all measured costs are subject to
change if you measure these costs on a different processor or under a
different system configuration. The key point is that no matter how well the
compiler and runtime are developed and tuned to minimize the overhead of
OpenMP constructs and clauses, you can always find ways to reduce the
overhead by exploring the use of OpenMP in a more effective way.

Table 6.4 Measured Cost of OpenMP Constructs and Clauses

Constructs Cost (in microseconds) Scalability
parallel 1.5 Linear
Barrier 1.0 Linear or O(log(n))
schedule(static) 1.0 Linear
schedule(guided) 6.0 Depends on contention
schedule(dynamic) 50 Depends on contention
ordered 0.5 Depends on contention
Single 1.0 Depends on contention
Reduction 2.5 Linear or O(log(n))
Atomic 0.5 Depends on data-type and
hardware
Critical 0.5 Depends on contention
Lock/Unlock 0.5 Depends on contention

Earlier, you saw how the parallel for pragma could be used to
split the iterations of a loop across multiple threads. When the compiler
generated thread is executed, the iterations of the loop are distributed
among threads. At the end of the parallel region, the threads are
suspended and they wait for the next parallel region, loop, or sections.
A suspend or resume operation, while significantly lighter weight than
create or terminate operations, still creates overhead and may be
unnecessary when two parallel regions, loops, or sections are adjacent as
shown in the following example.
#pragma omp parallel for for
( k = 0; k < m; k++ ) {
fn1(k); fn2(k);
}

#pragma omp parallel for // adds unnecessary overhead

for ( k = 0; k < m; k++ ) {
fn3(k); fn4(k);
}
 Chapter 6: OpenMP: A Portable Solution for Threading 151

The overhead can be removed by entering a parallel region once,

then dividing the work within the parallel region. The following code is
functionally identical to the preceding code but runs faster, because the
overhead of entering a parallel region is performed only once.
#pragma omp parallel
{
#pragma omp for
for ( k = 0; k < m; k++ ) {
fn1(k); fn2(k);
}

#pragma omp for

for ( k = 0; k < m; k++ ) {
fn3(k); fn4(k);
}
}
Ideally, all performance-critical portions of the application would be
executed within a parallel region. Since very few programs are
comprised only of loops, additional constructs are used to handle non-
loop code. A work-sharing section is one such construct.

Work-sharing Sections
The work-sharing sections construct directs the OpenMP compiler and
runtime to distribute the identified sections of your application among
threads in the team created for the parallel region. The following
example uses work-sharing for loops and work-sharing sections
together within a single parallel region. In this case, the overhead of
forking or resuming threads for parallel sections is eliminated.
#pragma omp parallel
{
#pragma omp for
for ( k = 0; k < m; k++ ) {
x = fn1(k) + fn2(k);
}

#pragma omp sections private(y, z)

{
#pragma omp section
{ y = sectionA(x); fn7(y); }
#pragma omp section
{ z = sectionB(x); fn8(z); }
}
}
152 Multi-Core Programming

Here, OpenMP first creates several threads. Then, the iterations of the
loop are divided among the threads. Once the loop is finished, the
sections are divided among the threads so that each section is executed
exactly once, but in parallel with the other sections. If the program
contains more sections than threads, the remaining sections get
scheduled as threads finish their previous sections. Unlike loop
scheduling, the schedule clause is not defined for sections. Therefore,
OpenMP is in complete control of how, when, and in what order threads
are scheduled to execute the sections. You can still control which
variables are shared or private, using the private and reduction
clauses in the same fashion as the loop construct.

Performance-oriented Programming
OpenMP provides a set of important pragmas and runtime functions that
enable thread synchronization and related actions to facilitate correct
parallel programming. Using these pragmas and runtime functions
effectively with minimum overhead and thread waiting time is extremely
important for achieving optimal performance from your applications.

Using Barrier and Nowait

Barriers are a form of synchronization method that OpenMP employs to
synchronize threads. Threads will wait at a barrier until all the threads in
the parallel region have reached the same point. You have been using
implied barriers without realizing it in the work-sharing for and work-
sharing sections constructs. At the end of the parallel, for,
sections, and single constructs, an implicit barrier is generated by the
compiler or invoked in the runtime library. The barrier causes execution to
wait for all threads to finish the work of the loop, sections, or region
before any go on to execute additional work. This barrier can be removed
with the nowait clause, as shown in the following code sample.
#pragma omp parallel
{
#pragma omp for nowait
for ( k = 0; k < m; k++ ) {
fn10(k); fn20(k);
}

#pragma omp sections private(y, z)

{
#pragma omp section
{ y = sectionD(); fn70(y); }
 Chapter 6: OpenMP: A Portable Solution for Threading 153

#pragma omp section

{ z = sectionC(); fn80(z); }
}
}
In this example, since data is not dependent between the first work-
sharing for loop and the second work-sharing sections code block, the
threads that process the first work-sharing for loop continue
immediately to the second work-sharing sections without waiting for
all threads to finish the first loop. Depending upon your situation, this
behavior may be beneficial, because it can make full use of available
resources and reduce the amount of time that threads are idle. The
nowait clause can also be used with the work-sharing sections
construct and single construct to remove its implicit barrier at the end
of the code block.
Adding an explicit barrier is also supported by OpenMP as shown in
the following example through the barrier pragma.
#pragma omp parallel shared(x, y, z) num_threads(2)
{
int tid = omp_get_thread_num();
if (tid == 0) {
y = fn70(tid);
}
else {
z = fn80(tid);
}

#pragma omp barrier

#pragma omp for
for ( k = 0; k < 100; k++ ) {
x[k] = y + z + fn10(k) + fn20(k);
}
}
In this example, the OpenMP code is to be executed by two threads;
one thread writes the result to the variable y, and another thread writes
the result to the variable z. Both y and z are read in the work-sharing for
loop, hence, two flow dependences exist. In order to obey the data
dependence constraints in the code for correct threading, you need to
add an explicit barrier pragma right before the work-sharing for loop
to guarantee that the value of both y and z are ready for read. In real
applications, the barrier pragma is especially useful when all threads
need to finish a task before any more work can be completed, as would
be the case, for example, when updating a graphics frame buffer before
displaying its contents.
154 Multi-Core Programming

Interleaving Single-thread and Multi-thread Execution

In large real-world applications, a program may consist of both serial and
parallel code segments due to various reasons such as data dependence
constraints and I/O operations. A need to execute something only once
by only one thread will certainly be required within a parallel region,
especially because you are making parallel regions as large as possible
to reduce overhead. To handle the need for single-thread execution,
OpenMP provides a way to specify that a sequence of code contained
within a parallel section should only be executed one time by only one
thread. The OpenMP runtime library decides which single thread will
do the execution. If need be, you can specify that you want only the
master thread, the thread that actually started the program execution, to
execute the code, as in the following example.
#pragma omp parallel
{ // every thread calls this function
int tid = omp_get_thread_num();

// this loop is divided among the threads

#pragma omp for nowait
for ( k = 0; k < 100; k++ ) x[k] = fn1(tid);
// no implicit barrier at the end of the above loop causes
// all threads to synchronize here

#pragma omp master

y = fn_input_only();// only the master thread calls this

// adding an explicit barrier to synchronize all threads

// to make sure x[0-99] and y is ready for use
#pragma omp barrier
// again, this loop is divided among the threads
#pragma omp for nowait
for ( k = 0; k < 100; k++ ) x[k] = y + fn2(x[k]);
// The above loop does not have an implicit barrier, so
// threads will not wait for each other.
// One thread presumbly the first one done with above --
// will continue and execute the following code.
#pragma omp single
fn_single_print(y); // only one of threads calls this
// The above single construct has an implicit barrier,
// so all threads synchronize here before printing x[].
#pragma omp master
fn_print_array(x); // only one of threads prints x[]
}
 Chapter 6: OpenMP: A Portable Solution for Threading 155

As can be seen from the comments in this code, a remarkable amount

of synchronization and management of thread execution is available in a
comparatively compact lexicon of pragmas. Note that all low-level details
are taken care of by the OpenMP compiler and runtime. What you need
to focus on is to specify parallel computation and synchronization
behaviors you expected for correctness and performance. In other
words, using single and master pragmas along with the barrier
pragma and nowait clause in a clever way, you should be able to
maximize the scope of a parallel region and the overlap of computations
to reduce threading overhead effectively, while obeying all data
dependences and I/O constraints in your programs.

Data Copy-in and Copy-out

When you parallelize a program, you would normally have to deal with how
to copy in the initial value of a private variable to initialize its private copy for
each thread in the team. You would also copy out the value of the private
variable computed in the last iteration/section to its original variable for the
master thread at the end of parallel region. OpenMP standard provides four
clausesfirstprivate, lastprivate, copyin, and copyprivatefor
you to accomplish the data copy-in and copy-out operations whenever
necessary based on your program and parallelization scheme. The following
descriptions summarize the semantics of these four clauses:
firstprivate provides a way to initialize the value of a private
variable for each thread with the value of variable from the master
thread. Normally, temporary private variables have an undefined
initial value saving the performance overhead of the copy.
lastprivate provides a way to copy out the value of the private
variable computed in the last iteration/section to the copy of the
variable in the master thread. Variables can be declared both
firstprivate and lastprivate at the same time.
copyin provides a way to copy the master threads
threadprivate variable to the threadprivate variable of each
other member of the team executing the parallel region.
copyprivate provides a way to use a private variable to
broadcast a value from one member of threads to other members
of the team executing the parallel region. The copyprivate
clause is allowed to associate with the single construct; the
broadcast action is completed before any of threads in the team
left the barrier at the end of construct.
156 Multi-Core Programming

Considering the code example, lets see how it works. The following
code converts a color image to black and white.
for ( row = 0; row < height; row++ ) {
for ( col = 0; col < width; col++ ) {
pGray[col] = (BYTE)
( pRGB[row].red * 0.299 +
pRGB[row].green * 0.587 +
pRGB[row].blue * 0.114 );
}
pGray += GrayStride;
pRGB += RGBStride;
}
The issue is how to move the pointers pGray and pRGB to the correct
place within the bitmap while threading the outer row loop. The
address computation for each pixel can be done with the following code:
pDestLoc = pGray + col + row * GrayStride;
pSrcLoc = pRGB + col + row * RGBStride;
The above code, however, executes extra math on each pixel for the
address computation. Instead, the firstprivate clause can be used to
perform necessary initialization to get the initial address of pointer pGray
and pRGB for each thread. You may notice that the initial addresses of the
pointer pGray and pRGB have to be computed only once based on the
row number and their initial addresses in the master thread for each
thread; the pointer pGray and pRGB are induction pointers and updated
in the outer loop for each row iteration. This is the reason the bool-
type variable doInit is introduced with an initial value TRUE to make
sure the initialization is done only once for each to compute the initial
address of pointer pGray and pRGB. The parallelized code follows:
#pragma omp parallel for private (row, col) \
firstprivate(doInit, pGray, pRGB)
for ( row = 0; row < height; row++ ) {
// Need this init test to be able to start at an
// arbitrary point within the image after threading.
if (doInit == TRUE) {
doInit = FALSE;
pRGB += ( row * RGBStride );
pGray += ( row * GrayStride );
}
for ( col = 0; col < width; col++ ) {
pGray[col] = (BYTE) ( pRGB[row].red * 0.299 +
pRGB[row].green * 0.587 +
pRGB[row].blue * 0.114 );
 Chapter 6: OpenMP: A Portable Solution for Threading 157

}
pGray += GrayStride;
pRGB += RGBStride;
}

If you take a close look at this code, you may find that the four variables
GrayStride, RGBStride, height, and width are read-only variables. In
other words, no write operation is performed to these variables in the
parallel loop. Thus, you can also specify them on the parallel for loop
by adding the code below:
firstprivate (GrayStride, RGBStride, height, width)
You may get better performance in some cases, as the privatization
helps the compiler to perform more aggressive registerization and code
motion as their loop invariants reduce memory traffic.

Protecting Updates of Shared Variables

The critical and atomic pragmas are supported by the OpenMP
standard for you to protect the updating of shared variables for avoiding
data-race conditions. The code block enclosed by a critical section and an
atomic pragma are areas of code that may be entered only when no other
thread is executing in them. The following example uses an unnamed
critical section.
#pragma omp critical
{
if ( max < new_value ) max = new_value
}
Global, or unnamed, critical sections will likely and unnecessarily
affect performance because every thread is competing to enter the
same global critical section, as the execution of every thread is
serialized. This is rarely what you want. For this reason, OpenMP offers
named critical sections. Named critical sections enable fine-grained
synchronization, so only the threads that need to block on a particular
section will do so. The following example shows the code that
improves the previous example. In practice, named critical sections are
used when more than one thread is competing for more than one
critical resource.
#pragma omp critical(maxvalue)
{
if ( max < new_value ) max = new_value
}
158 Multi-Core Programming

With named critical sections, applications can have multiple critical

sections, and threads can be in more than one critical section at a time. It is
important to remember that entering nested critical sections runs the risk
of deadlock. The following code example code shows a deadlock situation:
void dequeue(NODE *node)
{
#pragma omp critical (x)
{
node = node->next;
}
}

void do_work(NODE *node)

{
#pragma omp critical (x)
{
node->next->data = fn1(node->data);
node = dequeue(node)
}
}

In the previous code, the dynamically nested critical sections are used.
When the function do_work is called inside a parallel loop, multiple
threads compete to enter the outer critical section. The thread that
succeeds in entering the outer critical section will call the dequeue
function; however, the dequeue function cannot make any further
progress, as the inner critical section attempts to enter the same critical
section in the do_work function. Thus, the do_work function could
never complete. This is a deadlock situation. The simple way to fix the
problem in the previous code is to do the inlining of the dequeue
function in the do_work function as follows:
void do_work(NODE *node)
{
#pragma omp critical (x)
{
node->next->data = fn1(node->data);
node = node->next;
}
}
When using multiple critical sections, be very careful to examine
critical sections that might be lurking in subroutines. In addition to using
critical sections, you can also use the atomic pragma for updating
shared variables. When executing code in parallel, it is impossible to
know when an operation will be interrupted by the thread scheduler.
 Chapter 6: OpenMP: A Portable Solution for Threading 159

However, sometimes you may require that a statement in a high-level

language complete in its entirety before a thread is suspended. For
example, a statement x++ is translated into a sequence of machine
instructions such as:
load reg, [x];
add reg 1;
store [x], reg;

It is possible that the thread is swapped out between two of these machine
instructions. The atomic pragma directs the compiler to generate code to
ensure that the specific memory storage is updated atomically. The
following code example shows a usage of the atomic pragma.
int main()
{ float y[1000];
int k, idx[1000];

#pragma omp parallel for shared(y, idx)

for ( k = 0; k < 8000; k++) {
idx[k] = k % 1000;
y[idx[k]] = 8.0;
}

#pragma omp parallel for shared(y, idx)

for ( k = 0; k < 8000; k++) {
#pragma omp atomic
y[idx[k]] += 8.0 * (k % 1000);
}
return 0;
}
An expression statement that is allowed to use the atomic pragma
must be with one of the following forms:
x binop = expr
x++
++x
x --
-- x
In the preceding expressions, x is an lvalue expression with scalar type;
expr is an expression with scalar type and does not reference the object
designed by x; binop is not an overloaded operator and is one of +, *, -, /,
&, ^, |, <<, or >> for the C/C++ language.
It is worthwhile to point out that in the preceding code example,
the advantage of using the atomic pragma is that it allows update of
160 Multi-Core Programming

two different elements of array y to occur in parallel. If a critical section

were used instead, then all updates to elements of array y would be
executed serially, but not in a guaranteed order. Furthermore, in
general, the OpenMP compiler and runtime library select the most
efficient method to implement the atomic pragma given operating
system features and hardware capabilities. Thus, whenever it is possible
you should use the atomic pragma before using the critical section in
order to avoid data-race conditions on statements that update a shared
memory location.

Intel Taskqueuing Extension to OpenMP

The Intel Taskqueuing extension to OpenMP allows a programmer to
parallelize control structures such as recursive function, dynamic-tree
search, and pointer-chasing while loops that are beyond the scope of
those supported by the current OpenMP model, while still fitting into the
framework defined by the OpenMP specification. In particular, the
taskqueuing model is a flexible programming model for specifying units
of work that are not pre-computed at the start of the work-sharing
construct. Take a look the following example.
void tq_func(LIST *p)
{
#pragma intel omp parallel taskq shared(p)
{ while (p!= NULL) {
#pragma intel omp task captureprivate(p)
{ tq_work1(p, 70); }
#pragma intel omp task captureprivate(p)
{ tq_work2(p, 80); }
p= p->next;
}
}
}

The parallel taskq pragma directs the compiler to generate code to

create a team of threads and an environment for the while loop to
enqueue the units of work specified by the enclosed task pragma. The
loops control structure and the enqueuing are executed by one thread,
while the other threads in the team participate in dequeuing the work
from the taskq queue and executing it. The captureprivate clause
ensures that a private copy of the link pointer p is captured at the time
each task is being enqueued, hence preserving the sequential semantics.
The taskqueuing execution model is shown in Figure 6.1.
 Chapter 6: OpenMP: A Portable Solution for Threading 161

Enqueue taskq Enqueue task

Thread pool

T1 T2 ... TK ... TN

Tm (m=1 ... N, and m = K) Work queue

Dequeue task (work unit)

Work queue empty

Schedule task (work unit) Done

Figure 6.1 Taskqueuing Execution Model

Essentially, for any given program with parallel taskq constructs,

a team of threads is created by the runtime library when the main thread
encounters a parallel region. The runtime thread scheduler chooses one
thread TK to execute initially from all the threads that encounter a taskq
pragma. All the other threads wait for work to be put on the task queue.
Conceptually, the taskq pragma triggers this sequence of actions:
1. Causes an empty queue to be created by the chosen thread TK
2. Enqueues each task that it encounters
3. Executes the code inside the taskq block as a single thread
The task pragma specifies a unit of work, potentially to be executed by a
different thread. When a task pragma is encountered lexically within a
taskq block, the code inside the task block is placed on the queue
associated with the taskq pragma. The conceptual queue is disbanded
when all work enqueued on it finishes and the end of the taskq block is
reached. The Intel C++ compiler has been extended throughout its
various components to support the taskqueuing model for generating
multithreaded codes corresponding to taskqueuing constructs.
162 Multi-Core Programming

OpenMP Library Functions

As you may remember, in addition to pragmas, OpenMP provides a set of
functions calls and environment variables. So far, only the pragmas have
been described. The pragmas are the key to OpenMP because they provide
the highest degree of simplicity and portability, and the pragmas can be
easily switched off to generate a non-threaded version of the code.
In contrast, the OpenMP function calls require you to add the
conditional compilation in your programs as shown below, in case you
want to generate a serial version.
#include <omp.h>
#ifdef _OPENMP
omp_set_num_threads(4);
#endif
When in doubt, always try to use the pragmas and keep the function
calls for the times when they are absolutely necessary. To use the
function calls, include the <omp.h> header file. The compiler
automatically links to the correct libraries.
The four most heavily used OpenMP library functions are shown in
Table 6.5. They retrieve the total number of threads, set the number of
threads, return the current thread number, and return the number of
available cores, logical processors or physical processors, respectively.
To view the complete list of OpenMP library functions, please see the
OpenMP Specification Version 2.5, which is available from OpenMP web
site at www.openmp.org.
Table 6.5 The Most Heavily Used OpenMP Library Functions
Function Name Description
int omp_get_num_threads ( void ); Returns the number of threads currently in
use. If called outside a parallel region, this
function will return 1.
int omp_set_num_threads ( int This function sets the number of threads that
NumThreads ); will be used when entering a parallel section.
It overrides the OMP_NUM_THREADS
environment variable.
int omp_get_thread_num ( void ); Returns the current thread number between 0
(master thread) and total number of threads - 1.
int omp_get_num_procs ( void ); Returns the number of available cores (or
processors). A core or processor with Hyper-
Threading Technology enabled will count as
two cores (or two processors).
 Chapter 6: OpenMP: A Portable Solution for Threading 163

Figure 6.2 uses these functions to perform data processing for each
element in array x. This example illustrates a few important concepts
when using the function calls instead of pragmas. First, your code must
be rewritten, and with any rewrite comes extra documentation,
debugging, testing, and maintenance work. Second, it becomes difficult
or impossible to compile without OpenMP support. Finally, because
thread values have been hard coded, you lose the ability to have loop-
scheduling adjusted for you, and this threaded code is not scalable
beyond four cores or processors, even if you have more than four cores
or processors in the system.

float x[8000];
omp_set_num_threads(4);
#pragma omp parallel private(k)
{ // This code has a shortcoming. Can you find it?
int num_thds = omp_get_num_threads();
int ElementsPerThread = 8000 / num_thds;
int Tid = omp_get_thread_num();
int LowBound = Tid*ElementsPerThread;
int UpperBound = LowBound + ElementsPerThread;

for ( k = LowBound; k < UpperBound; k++ )

DataProcess(x[k]);
}

Figure 6.2 Loop that Uses OpenMP Functions and Illustrates the Drawbacks

OpenMP Environment Variables

The OpenMP specification defines a few environment variables. Occasionally
the two shown in Table 6.6 may be useful during development.

Table 6.6 Most Commonly Used Environment Variables for OpenMP

Environment Variable Description Example
OMP_SCHEDULE Controls the scheduling of set OMP_SCHEDULE = guided, 2
the for-loop work-sharing
construct.
OMP_NUM_THREADS Sets the default number set OMP_NUM_THREADS = 4
of threads. The
omp_set_num_threads()
function call can override
this value.
164 Multi-Core Programming

Additional compiler-specific environment variables are usually

available. Be sure to review your compilers documentation to become
familiar with additional variables.

Compilation
Using the OpenMP pragmas requires an OpenMP-compatible compiler
and thread-safe runtime libraries. The Intel C++ Compiler version 7.0
or later and the Intel Fortran compiler both support OpenMP on Linux
and Windows. This books discussion of compilation and debugging
will focus on these compilers. Several other choices are available as
well, for instance, Microsoft supports OpenMP in Visual C++ 2005 for
Windows and the Xbox 360 platform, and has also made OpenMP
work with managed C++ code. In addition, OpenMP compilers for
C/C++ and Fortran on Linux and Windows are available from the
Portland Group.
The /Qopenmp command-line option given to the Intel C++ Compiler
instructs it to pay attention to the OpenMP pragmas and to create
multithreaded code. If you omit this switch from the command line, the
compiler will ignore the OpenMP pragmas. This action provides a very
simple way to generate a single-threaded version without changing any
source code. Table 6.7 provides a summary of invocation options for
using OpenMP.

Table 6.7 Compiler Switches for OpenMP (C/C++ and Fortran)

Windows Linux Semantics

-Qopenmp -openmp Generate multithreaded code for Intel

Pentium III, Pentium 4 with Hyper-
Threading Technology, Pentium M,
and multi-core processors.
-Qopenmp-profile -openmp-profile Link with instrumented OpenMP
runtime library to generate OpenMP
profiling information for use with the
OpenMP component of VTune
Performance Analyzer.
-Qopenmp-stubs -openmp-stubs Enable the user to compile OpenMP
programs in sequential mode. The
openmp directives are ignored and a
stub OpenMP library is linked for
sequential execution.
 Chapter 6: OpenMP: A Portable Solution for Threading 165

Windows Linux Semantics

-Qopenmp-report -openmp-report Control level of reports:
0 - Disable parallelization diagnostics
1 - report successfully threaded code
0 0
[default]
1 1
2 1 + report successfully code
2 2
generation for master, single, critical,
and atomic.

For conditional compilation, the compiler defines _OPENMP. If needed,

this definition can be tested in this manner:
#ifdef _OPENMP
printf ( "Hello World, I'm using OpenMP!\n" );
#endif

The thread-safe runtime libraries are selected and linked automatically

when the OpenMP related compilation switch is used.
The Intel compilers support the OpenMP Specification Version 2.5
except the workshare construct. Be sure to browse the release notes
and compatibility information supplied with the compiler for the latest
information. The complete OpenMP specification is available from the
OpenMP Web site, listed in References. To review details about OpenMP
for the Intel Compiler, see Chapter 11.

Debugging
Debugging multithreaded applications has always been a challenge due
to the nondeterministic execution of multiple instruction streams caused
by runtime thread-scheduling and context switching. Also, debuggers
may change the runtime performance and thread scheduling behaviors,
which can mask race conditions and other forms of thread interaction.
Even print statements can mask issues because they use synchronization
and operating system functions to guarantee thread-safety.
Debugging an OpenMP program adds some difficulty, as OpenMP
compilers must communicate all the necessary information of private
variables, shared variables, threadprivate variables, and all kinds of
constructs to debuggers after threaded code generation; additional code
that is impossible to examine and step through without a specialized
OpenMP-aware debugger. Therefore, the key is narrowing down the
problem to a small code section that causes the same problem. It would
be even better if you could come up with a very small test case that can
166 Multi-Core Programming

reproduce the problem. The following list provides guidelines for

debugging OpenMP programs.
1. Use the binary search method to identify the parallel construct
causing the failure by enabling and disabling the OpenMP
pragmas in the program.
2. Compile the routine causing problem with no /Qopenmp switch
and with /Qopenmp_stubs switch; then you can check if the
code fails with a serial run, if so, it is a serial code debugging. If
not, go to Step 3.
3. Compile the routine causing problem with /Qopenmp switch and
set the environment variable OMP_NUM_THREADS=1; then you
can check if the threaded code fails with a serial run. If so, it is a
single-thread code debugging of threaded code. If not, go to Step 4.
4. Identify the failing scenario at the lowest compiler optimization
level by compiling it with /Qopenmp and one of the switches
such as /Od, /O1, /O2, /O3, and/or /Qipo.
5. Examine the code section causing the failure and look
for problems such as violation of data dependence after paralleliza-
tion, race conditions, deadlock, missing barriers, and uninitialized
variables. If you can not spot any problem, go to Step 6.
6. Compile the code using /Qtcheck to perform the OpenMP code
instrumentation and run the instrumented code inside the Intel
Thread Checker.
Problems are often due to race conditions. Most race conditions are
caused by shared variables that really should have been declared private,
reduction, or threadprivate. Sometimes, race conditions are also caused
by missing necessary synchronization such as critica and atomic
protection of updating shared variables. Start by looking at the variables
inside the parallel regions and make sure that the variables are declared
private when necessary. Also, check functions called within parallel
constructs. By default, variables declared on the stack are private but the
C/C++ keyword static changes the variable to be placed on the global
heap and therefore the variables are shared for OpenMP loops. The
default(none) clause, shown in the following code sample, can be
used to help find those hard-to-spot variables. If you specify
default(none), then every variable must be declared with a data-
sharing attribute clause.
#pragma omp parallel for default(none) private(x,y) shared(a,b)
 Chapter 6: OpenMP: A Portable Solution for Threading 167

Another common mistake is uninitialized variables. Remember that

private variables do not have initial values upon entering or exiting a
parallel construct. Use the firstprivate or lastprivate clauses
discussed previously to initialize or copy them. But do so only when
necessary because this copying adds overhead.
If you still cant find the bug, perhaps you are working with just too
much parallel code. It may be useful to make some sections execute
serially, by disabling the parallel code. This will at least identify the
location of the bug. An easy way to make a parallel region execute in
serial is to use the if clause, which can be added to any parallel
construct as shown in the following two examples.
#pragma omp parallel if(0)
printf("Executed by thread %d\n", omp_get_thread_num());

#pragma omp parallel for if(0)

for ( x = 0; x < 15; x++ ) fn1(x);

In the general form, the if clause can be any scalar expression, like the
one shown in the following example that causes serial execution when
the number of iterations is less than 16.
#pragma omp parallel for if(n>=16)
for ( k = 0; k < n; k++ ) fn2(k);

Another method is to pick the region of the code that contains the bug
and place it within a critical section, a single construct, or a master
construct. Try to find the section of code that suddenly works when it is
within a critical section and fails without the critical section, or executed
with a single thread.
The goal is to use the abilities of OpenMP to quickly shift code back
and forth between parallel and serial states so that you can identify the
locale of the bug. This approach only works if the program does in
fact function correctly when run completely in serial mode. Notice that
only OpenMP gives you the possibility of testing code this way without
rewriting it substantially. Standard programming techniques used in the
Windows API or Pthreads irretrievably commit the code to a threaded
model and so make this debugging approach more difficult.

Performance
OpenMP paves a simple and portable way for you to parallelize your
applications or to develop threaded applications. The threaded application
168 Multi-Core Programming

performance with OpenMP is largely dependent upon the following

factors:
The underlying performance of the single-threaded code.
The percentage of the program that is run in parallel and its
scalability.
CPU utilization, effective data sharing, data locality and load
balancing.
The amount of synchronization and communication among the
threads.
The overhead introduced to create, resume, manage, suspend,
destroy, and synchronize the threads, and made worse by the
number of serial-to-parallel or parallel-to-serial transitions.
Memory conflicts caused by shared memory or falsely shared
memory.
Performance limitations of shared resources such as memory, write
combining buffers, bus bandwidth, and CPU execution units.
Essentially, threaded code performance boils down to two issues: how
well does the single-threaded version run, and how well can the work be
divided up among multiple processors with the least amount of
overhead?
Performance always begins with a well-designed parallel algorithm or
well-tuned application. The wrong algorithm, even one written in hand-
optimized assembly language, is just not a good place to start. Creating a
program that runs well on two cores or processors is not as desirable as
creating one that runs well on any number of cores or processors.
Remember, by default, with OpenMP the number of threads is chosen by
the compiler and runtime librarynot youso programs that work well
regardless of the number of threads are far more desirable.
Once the algorithm is in place, it is time to make sure that the code
runs efficiently on the Intel Architecture and a single-threaded version
can be a big help. By turning off the OpenMP compiler option you can
generate a single-threaded version and run it through the usual set of
optimizations. A good reference for optimizations is The Software
Optimization Cookbook (Gerber 2006). Once you have gotten the
single-threaded performance that you desire, then it is time to generate
the multithreaded version and start doing some analysis.
 Chapter 6: OpenMP: A Portable Solution for Threading 169

First look at the amount of time spent in the operating systems idle
loop. The Intel VTune Performance Analyzer is great tool to help with
the investigation. Idle time can indicate unbalanced loads, lots of blocked
synchronization, and serial regions. Fix those issues, then go back to the
VTune Performance Analyzer to look for excessive cache misses and
memory issues like false-sharing. Solve these basic problems, and you will
have a well-optimized parallel program that will run well on multi-core
systems as well as multiprocessor SMP systems.
Optimizations are really a combination of patience, trial and error, and
practice. Make little test programs that mimic the way your application
uses the computers resources to get a feel for what things are faster than
others. Be sure to try the different scheduling clauses for the parallel
sections. Chapter 7 provides additional advice on how to tune parallel
code for performance and Chapter 11 covers the tools youll need.

Key Points
Keep the following key points in mind while programming with OpenMP:
The OpenMP programming model provides an easy and portable
way to parallelize serial code with an OpenMP-compliant compiler.
OpenMP consists of a rich set of pragmas, environment variables,
and a runtime API for threading.
The environment variables and APIs should be used sparingly
because they can affect performance detrimentally. The pragmas
represent the real added value of OpenMP.
With the rich set of OpenMP pragmas, you can incrementally
parallelize loops and straight-line code blocks such as sections
without re-architecting the applications. The Intel Task queuing
extension makes OpenMP even more powerful in covering more
application domain for threading.
If your applications performance is saturating a core or
processor, threading it with OpenMP will almost certainly
increase the applications performance on a multi-core or
multiprocessor system.
You can easily use pragmas and clauses to create critical sections,
identify private and public variables, copy variable values, and
control the number of threads operating in one section.
170 Multi-Core Programming

OpenMP automatically uses an appropriate number of threads for

the target system so, where possible, developers should consider
using OpenMP to ease their transition to parallel code and to make
their programs more portable and simpler to maintain. Native and
quasi-native options, such as the Windows threading API and
Pthreads, should be considered only when this is not possible.
 Chapter 7
Solutions to
Common Parallel
Programming
Problems
arallel programming has been around for decades, though before the
P advent of multi-core processors, it was an esoteric discipline.
Numerous programmers have tripped over the common stumbling
blocks by now. By recognizing these problems you can avoid stumbling.
Furthermore, it is important to understand the common problems before
designing a parallel program, because many of the problems arise from
the overall decomposition of the program, and cannot be easily patched
later. This chapter surveys some of these common problems, their
symptoms, and ways to circumvent them.

Too Many Threads

It may seem that if a little threading is good, then a lot must be better. In
fact, having too many threads can seriously degrade program perform-
ance. The impact comes in two ways. First, partitioning a fixed amount
of work among too many threads gives each thread too little work, so
that the overhead of starting and terminating threads swamps the useful
work. Second, having too many concurrent software threads incurs
overhead from having to share fixed hardware resources.
When there are more software threads than hardware threads, the
operating system typically resorts to round robin scheduling. The

171
172 Multi-Core Programming

scheduler gives each software thread a short turn, called a time slice, to
run on one of the hardware threads. When a software threads time slice
runs out, the scheduler preemptively suspends the thread in order to run
another software thread on the same hardware thread. The software
thread freezes in time until it gets another time slice.
Time slicing ensures that all software threads make some progress.
Otherwise, some software threads might hog all the hardware threads
and starve other software threads. However, this equitable distribution of
hardware threads incurs overhead. When there are too many software
threads, the overhead can severely degrade performance. There are
several kinds of overhead, and it helps to know the culprits so you can
spot them when they appear.
The most obvious overhead is the process of saving and restoring a
threads register state. Suspending a software thread requires saving the
register values of the hardware thread, so the values can be restored
later, when the software thread resumes on its next time slice. Typically,
thread schedulers allocate big enough time slices so that the save/restore
overheads for registers are insignificant, so this obvious overhead is in
fact not much of a concern.
A more subtle overhead of time slicing is saving and restoring a
threads cache state. Modern processors rely heavily on cache
memory, which can be about 10 to 100 times faster than main
memory. Accesses that hit in cache are not only much faster; they also
consume no bandwidth of the memory bus. Caches are fast, but finite.
When the cache is full, a processor must evict data from the cache to
make room for new data. Typically, the choice for eviction is the least
recently used data, which more often than not is data from an earlier
time slice. Thus threads tend to evict each others data. The net effect
is that too many threads hurt performance by fighting each other for
cache.
A similar overhead, at a different level, is thrashing virtual memory.
Most systems use virtual memory, where the processors have an address
space bigger than the actual available memory. Virtual memory resides
on disk, and the frequently used portions are kept in real memory.
Similar to caches, the least recently used data is evicted from memory
when necessary to make room. Each software thread requires virtual
memory for its stack and private data structures. As with caches, time
slicing causes threads to fight each other for real memory and thus hurts
performance. In extreme cases, there can be so many threads that the
program runs out of even virtual memory.
 Chapter 7: Solutions to Common Parallel Programming Problems 173

The cache and virtual memory issues described arise from sharing
limited resources among too many software threads. A very different, and
often more severe, problem arises called convoying, in which software
threads pile up waiting to acquire a lock. Consider what happens when a
threads time slice expires while the thread is holding a lock. All threads
waiting for the lock must now wait for the holding thread to wake up
and release the lock. The problem is even worse if the lock
implementation is fair, in which the lock is acquired in first-come first-
served order. If a waiting thread is suspended, then all threads waiting
behind it are blocked from acquiring the lock.
The solution that usually works best is to limit the number of
runnable threads to the number of hardware threads, and possibly
limit it to the number of outer-level caches. For example, a dual-core

Intel Pentium Processor Extreme Edition has two physical cores, each
with Hyper-Threading Technology, and each with its own cache. This
configuration supports four hardware threads and two outer-level
caches. Using all four runnable threads will work best unless the
threads need so much cache that it causes fighting over cache, in which
case maybe only two threads is best. The only way to be sure is to
experiment. Never hard code the number of threads; leave it as a
tuning parameter.
Runnable threads, not blocked threads, cause time-slicing overhead.
When a thread is blocked waiting for an external event, such as a mouse
click or disk I/O request, the operating system takes it off the round-
robin schedule. Hence a blocked thread does not cause time-slicing
overhead. A program may have many more software threads than
hardware threads, and still run efficiently if most of the OS threads are
blocked.
A helpful organizing principle is to separate compute threads from
I/O threads. Compute threads should be the threads that are runnable
most of the time. Ideally, the compute threads never block on external
events, and instead feed from task queues that provide work. The
number of compute threads should match the processor resources. The
I/O threads are threads that wait on external events most of the time, and
thus do not contribute to having too many threads.
Because building efficient task queues takes some expertise, it is
usually best to use existing software to do this. Common useful practices
are as follows:
Let OpenMP do the work. OpenMP lets the programmer specify
loop iterations instead of threads. OpenMP deals with managing
174 Multi-Core Programming

the threads. As long as the programmer does not request a

particular number of threads, the OpenMP implementation will
strive to use the optimal number of software threads.
Use a thread pool, which is a construct used to maintain a set of
long lived software threads and eliminates the overhead of
initialization process of threads for short lived tasks. A thread
pool is a collection of tasks which are serviced by the software
threads in the pool. Each software thread finishes a task before
taking on another. For example, Windows has a routine
QueueUserWorkItem. Clients add tasks by calling
QueueUserWorkItem with a callback and pointer that define the
task. Hardware threads feed from this queue. For managed code,
Windows .NET has a class ThreadPool. Java has a class Executor
for similar purposes. Unfortunately, there is no standard thread
pool support in POSIX threads.
Experts may wish to write their own task scheduler. The
method of choice is called work stealing, where each thread
has its own private collection of tasks. When a thread runs out
of tasks, it steals from another threads collection. Work
stealing yields good cache usage and load balancing. While a
thread is working on its own tasks, it tends to be reusing data
that is hot in its cache. When it runs out of tasks and has to
steal work, it balances the load. The trick to effective task
stealing is to bias the stealing towards large tasks, so that the
thief can stay busy for a while. The early Cilk scheduler
(Blumofe 1995) is a good example of how to write an effective
task-stealing scheduler.

Data Races, Deadlocks, and Live Locks

Unsynchronized access to shared memory can introduce race
conditions, where the program results depend nondeterministically on
the relative timings of two or more threads. Figure 7.1 shows two
threads trying to add to a shared variable x, which has an initial value of
0. Depending upon the relative speeds of the threads, the final value of
x can be 1, 2, or 3.
 Chapter 7: Solutions to Common Parallel Programming Problems 175

Figure 7.1 Unsynchronized Threads Racing against each Other Lead to

Nondeterministic Outcome

Parallel programming would be a lot easier if races were as obvious as

in Figure 7.1. But the same race can be hidden by language syntax in a
variety of ways, as shown by the examples in Figure 7.2. Update
176 Multi-Core Programming

operations such as += are normally just shorthand for temp = x; x =

temp+1, and hence can result in interleaving. Sometimes the shared
location is accessed by different expressions. Sometimes the shared
location is hidden by function calls. Even if each thread uses a single
instruction to fetch and update the location, there could be interleaving,
because the hardware might break the instruction into interleaved reads
and writes.

Figure 7.2 Race Conditions Hiding behind Language Syntax

Intel Thread Checker is a powerful tool for detecting potential race

conditions. It can see past all the varieties of camouflage shown in Figure 7.2
because it deals in terms of actual memory locations, not their names or
addressing expressions. Chapter 11 says more about Thread Checker.
Sometimes deliberate race conditions are intended and useful. For
example, threads may be reading a location that is updated
asynchronously with a latest current value. In such a situation, care
must be taken that the writes and reads are atomic. Otherwise, garbled
data may be written or read. For example, reads and writes of structure
types are often done a word at a time or a field at a time. Types longer
than the natural word size, such as 80-bit floating-point, might not be
read or written atomically, depending on the architecture. Likewise,
misaligned loads and stores, when supported, are usually not atomic. If
 Chapter 7: Solutions to Common Parallel Programming Problems 177

such an access straddles a cache line, the processor performs the access
as two separate accesses to the two constituent cache lines.
Data races can arise not only from unsynchronized access to shared
memory, but also from synchronized access that was synchronized at too
low a level. Figure 7.3 shows such an example. The intent is to use a list
to represent a set of keys. Each key should be in the list at most once.
Even if the individual list operations have safeguards against races, the
combination suffers a higher level race. If two threads both attempt to
insert the same key at the same time, they may simultaneously determine
that the key is not in the list, and then both would insert the key. What is
needed is a lock that protects not just the list, but that also protects the
invariant no key occurs twice in list.

Figure 7.3 A Higher-Level Race Condition Example.

Adding the necessary lock to correct Figure 7.3 exposes the

frustrating performance problem of locks. Building locks into low-level
components is often a waste of time, because the high-level components
that use the components will need higher-level locks anyway. The lower-
level locks then become pointless overhead. Fortunately, in such a
scenario the high-level locking causes the low-level locks to be
uncontended, and most lock implementations optimize the uncontended
case. Hence the performance impact is somewhat mitigated, but for best
performance the superfluous locks should be removed. Of course there
are times when components should provide their own internal locking.
This topic is discussed later in the discussion of thread-safe libraries.

Deadlock
Race conditions are typically cured by adding a lock that protects the
invariant that might otherwise be violated by interleaved operations.
Unfortunately, locks have their own hazards, most notably deadlock.
Figure 7.4 shows a deadlock involving two threads. Thread 1 has
acquired lock A. Thread 2 has acquired lock B. Each thread is trying to
acquire the other lock. Neither thread can proceed.
178 Multi-Core Programming

Figure 7.4 Deadlock Caused by Cycle

Though deadlock is often associated with locks, it can happen any

time a thread tries to acquire exclusive access to two more shared
resources. For example, the locks in Figure 7.4 could be files instead,
where the threads are trying to acquire exclusive file access.
Deadlock can occur only if the following four conditions hold true:
1. Access to each resource is exclusive.
2. A thread is allowed to hold one resource while requesting
another.
3. No thread is willing to relinquish a resource that it has acquired.
4. There is a cycle of threads trying to acquire resources, where
each resource is held by one thread and requested by another.
Deadlock can be avoided by breaking any one of these conditions.
Often the best way to avoid deadlock is to replicate a resource that
requires exclusive access, so that each thread can have its own private
copy. Each thread can access its own copy without needing a lock. The
copies can be merged into a single shared copy of the resource at the
end if necessary. By eliminating locking, replication avoids deadlock and
 Chapter 7: Solutions to Common Parallel Programming Problems 179

has the further benefit of possibly improving scalability, because the lock
that was removed might have been a source of contention.
If replication cannot be done, that is, in such cases where there
really must be only a single copy of the resource, common wisdom is to
always acquire the resources (locks) in the same order. Consistently
ordering acquisition prevents deadlock cycles. For instance, the
deadlock in Figure 7.4 cannot occur if threads always acquire lock A
before they acquire lock B.
The ordering rules that are most convenient depend upon the
specific situation. If the locks all have associated names, even something
as simple as alphabetical order works. This order may sound silly, but it
has been successfully used on at least one large project.
For multiple locks in a data structure, the order is often based on the
topology of the structure. In a linked list, for instance, the agreed upon
order might be to lock items in the order they appear in the list. In a tree
structure, the order might be a pre-order traversal of the tree. Somewhat
similarly, components often have a nested structure, where bigger
components are built from smaller components. For components nested
that way, a common order is to acquire locks in order from the outside to
the inside.
If there is no obvious ordering of locks, a solution is to sort the
locks by address. This approach requires that a thread know all locks
that it needs to acquire before it acquires any of them. For instance,
perhaps a thread needs to swap two containers pointed to by pointers x
and y, and each container is protected by a lock. The thread could
compare x < y to determine which container comes first, and acquire
the lock on the first container before acquiring a lock on the second
container, as Figure 7.5 illustrates.

void AcquireTwoLocksViaOrdering( Lock& x, Lock& y ) {

assert( &x!=&y );
if( &x<&y ) {
acquire x
acquire y
} else {
acquire y
acquire x
}
}

Figure 7.5 Locks Ordered by their Addresses

180 Multi-Core Programming

In large software projects, different programmers construct different

components, and by necessity should not have to understand the inner
workings of the other components. It follows that to prevent accidental
deadlock, software components should try to avoid holding a lock while
calling code outside the component, because the call chain may cycle
around and create a deadlock cycle.
The third condition for deadlock is that no thread is willing to give
up its claim on a resource. Thus another way of preventing deadlock is
for a thread to give up its claim on a resource if it cannot acquire the
other resources. For this purpose, mutexes often have some kind of try
lock routine that allows a thread to attempt to acquire a lock, and give
up if it cannot be acquired. This approach is useful in scenarios where
sorting the locks is impractical. Figure 7.6 sketches the logic for using a
try lock approach to acquire two locks, A and B. In Figure 7.6, a
thread tries to acquire both locks, and if it cannot, it release both locks
and tries again.

void AcquireTwoLocksViaBackoff( Lock& x, Lock& y ) {

for( int t=1; ; t*=2 ) {
acquire x
try to acquire y
if( y was acquired ) break;
release x
wait for random amount of time between 0 and t
}
}

Figure 7.6 Try and Back Off Logic

Figure 7.6 has some timing delays in it to prevent the hazard of live
lock. Live lock occurs when threads continually conflict with each other
and back off. Figure 7.6 applies exponential backoff to avoid live lock. If
a thread cannot acquire all the locks that it needs, it releases any that it
acquired and waits for a random amount of time. The random time is
chosen from an interval that doubles each time the thread backs off.
Eventually, the threads involved in the conflict will back off sufficiently
that at least one will make progress. The disadvantage of backoff schemes
is that they are not fair. There is no guarantee that a particular thread will
make progress. If fairness is an issue, then it is probably best to use lock
ordering to prevent deadlock.
 Chapter 7: Solutions to Common Parallel Programming Problems 181

Heavily Contended Locks

Proper use of lock to avoid race conditions can invite performance
problems if the lock becomes highly contended. The lock becomes like a
tollgate on a highway. If cars arrive at the tollgate faster than the toll
taker can process them, the cars will queue up in a traffic jam behind the
tollgate. Similarly, if threads try to acquire a lock faster than the rate at
which a thread can execute the corresponding critical section, then
program performance will suffer as threads will form a convoy waiting
to acquire the lock. Indeed, this behavior is sometimes referred to as
convoying.
As mentioned in the discussion of time-slicing woes, convoying
becomes even worse for fair locks, because if a thread falls asleep, all
threads behind it have to wait for it to wake up. Imagine that software
threads are cars and hardware threads are the drivers in those cars. This
might seem like a backwards analogy, but from a cars perspective,
people exist solely to move cars between parking places. If the cars form
a convoy, and a driver leaves his or her car, everyone else behind is
stuck.

Priority Inversion
Some threading implementations allow threads to have priorities. When
there are not enough hardware threads to run all software threads, the
higher priority software threads get preference. For example,
foreground tasks might be running with higher priorities than
background tasks. Priorities can be useful, but paradoxically, can lead
to situations where a low-priority thread blocks a high-priority thread
from running.
Figure 7.7 illustrates priority inversion. Continuing our analogy with
software threads as cars and hardware threads as drivers, three cars are
shown, but there is only a single driver. A low-priority car has acquired a
lock so it can cross a single-lane critical section bridge. Behind it waits
a high-priority car. But because the high-priority car is blocked, the driver
is attending the highest-priority runnable car, which is the medium-
priority one. As contrived as this sounds, it actually happened on the
NASA Mars Pathfinder mission.
182 Multi-Core Programming

Medium Priority Software Thread

High Priority Software Thread Low Priority Software Thread

Figure 7.7 Priority Inversion Scenario, Where High Priority Gets Blocked and
Medium Priority Gets the Cycles

In real life, the problem in Figure 7.7 would be solved by bumping up

the priority of the blocking car until it is out of the way. With locks, this
is called priority inheritance. When a high-priority thread needs to
acquire a lock held by a low-priority thread, the scheduler bumps up the
priority of the blocking thread until the lock is released. Indeed, the Mars
Pathfinder problem was solved by turning on priority inheritance (Reeves
1998).
An alternative is priority ceilings in which a priority, called the
ceiling, is assigned to the mutex. The ceiling is the highest priority of any
thread that is expected to hold the mutex. When a thread acquires the
mutex, its priority is immediately bumped up to the ceiling value for the
duration that it holds the mutex. The priority ceilings scheme is eager to
bump up a threads priority. In contrast, the priority inheritance scheme
is lazy by not bumping up a threads priority unless necessary.
Windows mutexes support priority inheritance by default. Pthreads
mutexes support neither the priority inheritance nor priority ceiling
protocols. Both protocols are optional in the pthreads standard. If they
exist in a particular implementation, they can be set for a mutex via the
function pthread_mutexattr_setprotocol and inspected with the
function pthread_mutexattr_getprotocol. Read the manual pages on
these functions to learn whether they are supported for the target
system.
 Chapter 7: Solutions to Common Parallel Programming Problems 183

Programmers rolling their own locks or busy waits may encounter

priority inversion if threads with different priorities are allowed to
acquire the same lock. Hand-coded spin locks are a common example. If
neither priority inheritance nor priority ceilings can be built into the lock
or busy wait, then it is probably best to restrict the locks contenders to
threads with the same priority.

Solutions for Heavily Contended Locks

Upon encountering a heavily contended lock, the first reaction of many
programmers is I need a faster lock. Indeed, some implementations of
locks are notoriously slow, and faster locks are possible. However, no
matter how fast the lock is, it must inherently serialize threads. A faster
lock can thus help performance by a constant factor, but will never
improve scalability. To improve scalability, either eliminate the lock or
spread out the contention.
The earlier discussion of deadlock mentioned the technique of
eliminating a lock by replicating the resource. That is certainly the
method of choice to eliminate lock contention if it is workable. For
example, consider contention for a counter of events. If each thread can
have its own private counter, then no lock is necessary. If the total count
is required, the counts can be summed after all threads are done
counting.
If the lock on a resource cannot be eliminated, consider partitioning
the resource and using a separate lock to protect each partition. The
partitioning can spread out contention among the locks. For example,
consider a hash table implementation where multiple threads might try
to do insertions at the same time. A simple approach to avoid race
conditions is to use a single lock to protect the entire table. The lock
allows only one thread into the table at a time. The drawback of this
approach is that all threads must contend for the same lock, which could
become a serial bottleneck. An alternative approach is to create an array
of sub-tables, each with its own lock, as shown in Figure 7.8. Keys can be
mapped to the sub-tables via a hashing function. For a given key, a thread
can figure out which table to inspect by using a hash function that
returns a sub-table index. Insertion of a key commences by hashing the
key to one of the sub-tables, and then doing the insertion on that sub-
table while holding the sub-tables lock. Given enough sub-tables and a
good hash function, the threads will mostly not contend for the same
sub-table and thus not contend for the same lock.
184 Multi-Core Programming

Figure 7.8 Spreading out Contention by Partitioning a Hash Table into

Multiple Sub-tables

Pursuit of the idea of spreading contention among multiple locks

further leads to fine-grained locking. For example, hash tables are
commonly implemented as an array of buckets, where each bucket holds
keys that hashed to the same array element. In fine-grained locking, there
might be a lock on each bucket. This way multiple threads can
concurrently access different buckets in parallel. This is straightforward
to implement if the number of buckets is fixed. If the number of buckets
has to be grown, the problem becomes more complicated, because
resizing the array may require excluding all but the thread doing the
resizing. A reader-writer lock helps solve this problem, as will be explained
shortly. Another pitfall is that if the buckets are very small, the space
overhead of the lock may dominate.
If a data structure is frequently read, but infrequently written, then a
reader-writer lock may help deal with contention. A reader-write lock
distinguishes readers from writers. Multiple readers can acquire the lock
at the same time, but only one writer can acquire it at a time. Readers
cannot acquire the lock while a writer holds it and vice-versa. Thus
readers contend only with writers.
The earlier fine-grained hash table is a good example of where
reader-write locks can help if the array of buckets must be dynamically
resizable. Figure 7.9 shows a possible implementation. The table
consists of an array descriptor that specifies the arrays size and
location. A reader-writer mutex protects this structure. Each bucket has
its own plain mutex protecting it. To access a bucket, a thread acquires
two locks: a reader lock on the array descriptor, and a lock on the
 Chapter 7: Solutions to Common Parallel Programming Problems 185

buckets mutex. The thread acquires a reader lock, not a writer lock, on
the reader-writer mutex even if it is planning to modify a bucket,
because the reader-writer mutex protects the array descriptor, not the
buckets. If a thread needs to resize the array, it requests a writer lock
on the reader-writer mutex. Once granted, the thread can safely modify
the array descriptor without introducing a race condition. The overall
advantage is that during times when the array is not being resized,
multiple threads accessing different buckets can proceed concurrently.
The principle disadvantage is that a thread must obtain two locks
instead of one. This increase in locking overhead can overwhelm the
advantages of increased concurrency if the table is typically not subject
to contention.

Figure 7.9 Hash Table with Fine-grained Locking

If writers are infrequent, reader-writer locks can greatly reduce

contention. However, reader-writer locks have limitations. When the rate
of incoming readers is very high, the lock implementation may suffer
from memory contention problems. Thus reader-writer locks can be very
useful for medium contention of readers, but may not be able to fix
problems with high contention. The reliable way to deal with high
contention is to rework the parallel decomposition in a way that lowers
the contention. For example, the schemes in Figures 7.8 and 7.9 might
be combined, so that a hash table is represented by a fixed number of
sub-tables, each with fine-grained locking.
186 Multi-Core Programming

Non-blocking Algorithms
One way to solve the problems introduced by locks is to not use locks.
Algorithms designed to do this are called non-blocking. The defining
characteristic of a non-blocking algorithm is that stopping a thread does
not prevent the rest of the system from making progress. There are
different non-blocking guarantees:
Obstruction freedom. A thread makes progress as long as there is
no contention, but live lock is possible. Exponential backoff can
be used to work around live lock.
Lock freedom. The system as a whole makes progress.
Wait freedom. Every thread makes progress, even when faced
with contention. Very few non-blocking algorithms achieve this.
Non-blocking algorithms are immune from lock contention, priority
inversion, and convoying. Non-blocking algorithms have a lot of
advantages, but with these come a new set of problems that need to be
understood.
Non-blocking algorithms are based on atomic operations, such as the
methods of the Interlocked class discussed in Chapter 5. A few non-
blocking algorithms are simple. Most are complex, because the
algorithms must handle all possible interleaving of instruction streams
from contending processors.
A trivial non-blocking algorithm is counting via an interlocked
increment instead of a lock. The interlocked instruction avoids lock
overhead and pathologies. However, simply using atomic operations is
not enough to avoid race conditions, because as discussed before,
composing thread-safe operations does not necessarily yield a thread-safe
procedure. As an example, the C code in Figure 7.10 shows the wrong
way and right way to decrement and test a reference count
p->ref_count. In the wrong code, if the count was originally 2, two
threads executing the wrong code might both decrement the count, and
then both see it as zero at the same time. The correct code performs the
decrement and test as a single atomic operation.
 Chapter 7: Solutions to Common Parallel Programming Problems 187

Figure 7.10 Atomic Operations and Race Conditions

Most non-blocking algorithms involve a loop that attempts to perform

an action using one or more compare-and-swap (CAS) operations, and
retries when one of the CAS operations fails. A simple and useful
example is implementing a thread-safe fetch-and-op. A fetch-and-op reads
a value from a location, computes a new value from it, and stores the
new value. Figure 7.11 illustrates both a locked version and a non-
blocking version that operate on a location x. The non-blocking version
reads location x into a local temporary x_old, and computes a new value
x_new = op(x_old). The routine InterlockedCompareExchange
stores the new value, unless x is now different than x_old. If the store
fails, the code starts over until it succeeds.

void LockedFetchAndOp( long& x ) {

acquire lock
x = op(x);
release lock
}

void NonBlockingFetchAndOp( volatile long& x ) {

long x_old, x_new, x_was;
do {
x_old = x;
x_new = op(x_old);
x_was = InterlockedCompareExchange(&x, x_new,
x_old);
} while( x_was!=x_old );
}

Figure 7.11 Comparison of Locked and Lockless Code for Fetch-and-op

188 Multi-Core Programming

Fetch-and-op is useful as long as the order in which various threads

perform op does not matter. For example, op might be multiply by 2.
The location x must have a type for which a compare-and-exchange
instruction is available.

ABA Problem
In Figure 7.11, there is a time interval between when a thread executes
x_old = x and when the thread executes InterlockedCompareEx-
change. During this interval, other processors might perform other fetch-
and-op operations. For example, suppose the initial value read is A. An
intervening sequence of fetch-and-op operations by other processors
might change x to B and then back to A. When the original thread
executes InterlockedCompareExchange, it will be as if the other
processors actions never happened. As long as the order in which op is
executed does not matter, there is no problem. The net result is the same
as if the fetch-and-op operations were reordered such that the
intervening sequence happens before the first read.
But sometimes fetch-and-op has uses where changing x from A to B
to A does make a difference. The problem is indeed known as the ABA
problem. Consider the lockless implementation of a stack shown in
Figure 7.12. It is written in the fetch-and-op style, and thus has the
advantage of not requiring any locks. But the op is no longer a pure
function, because it deals with another shared memory location: the field
next. Figure 7.13 shows a sequence where the function
BrokenLockLessPop corrupts the linked stack. When thread 1 starts
out, it sees B as next on stack. But intervening pushes and pops make C
next on stack. But Thread 1s final InterlockedCompareExchange does
not catch this switch because it only examines Top.

Node* Top; // Pointer to top item on stack.

void BrokenLocklessPush( Node* node ) {

Item *t_old, t_was;
do {
Item* t_old = Top;
n->next = t_old;
t_was = InterlockedCompareExchange(&Top,node,t_old);
} while( t_was!=t_old );
}
 Chapter 7: Solutions to Common Parallel Programming Problems 189

Item* BrokenLocklessPop() {
Item *t_old, *t_was, *t_new;
do {
t_old = Top;
t_new = t_old->next;
// ABA problem may strike below!
t_was = InterlockedCompareExchange(&Top,t_new,t_old);
} while( t_was!=t_old );
return t_old;
}

Figure 7.12 Lockless Implementation of a Linked Stack that May Suffer from ABA
Problem

Figure 7.13 Sequence Illustrates ABA Problem for Code in Figure 7.12

The solution to the ABA problem is to never reuse A. In a garbage-

collected environment such as Java or .NET, this is simply a matter of not
recycling nodes. That is, once a node has been popped, never push it
again. Instead allocate a fresh node. The garbage collector will do the
hard work of checking that the memory for node A is not recycled until
all extant references to it are gone.
190 Multi-Core Programming

In languages with garbage collection, the problem is harder. An old

technique dating back to the IBM 370 changes ABA to ABA. In other
words, make A slightly different each time. This is typically done by
appending a serial number to the pointer. A special instruction that can do
a double-wide compare-exchange is required. On IA-32, the instruction
is cmpxchg8b, which does a compare-exchange on eight bytes. On

processors with Intel EM64T, it is cmpxchg16b. On Itanium processors,
there is cmp8xchg16, which is not quite the same, because it compares
only the first eight bytes, but exchanges all 16. However, as long as the
serial number is the first eight bytes, it works for turning ABA into ABA.
Another solution is to build a miniature garbage collector that
handles pointers involved in compare-exchange operations. These
pointers are called hazard pointers, because they present a hazard to
lockless algorithms. Maged Michaels paper on hazard pointers (Michael
2004) explains how to implement hazard pointers. Hazard pointers are a
nontrivial exercise and make assumptions about the environment, so
tread with caution.

Cache Line Ping-ponging

Non-blocking algorithms can cause a lot of traffic on the memory bus as
various hardware threads keep trying and retrying to perform operations
on the same cache line. To service these operations, the cache line
bounces back and forth (ping-pongs) between the contending threads.
A locked algorithm may outperform the non-blocking equivalent if lock
contention is sufficiently distributed and each lock says hand off my
cache line until Im done. Experimentation is necessary to find out
whether the non-blocking or locked algorithm is better. A rough guide is
that a fast spin lock protecting a critical section with no atomic
operations may outperform an alternative non-blocking design that
requires three or more highly contended atomic operations.

Memory Reclamation Problem

Memory reclamation is the dirty laundry of many non-blocking
algorithms. For languages such as C/C++ that require the programmer to
explicitly free memory, it turns out to be surprisingly difficult to call
free on a node used in a non-blocking algorithm. Programmers planning
to use non-blocking algorithms need to understand when this limitation
arises, and how to work around it.
 Chapter 7: Solutions to Common Parallel Programming Problems 191

The problem occurs for algorithms that remove nodes from linked
structures, and do so by performing compare-exchange operations on
fields in the nodes. For example, non-blocking algorithms for queues do
this. The reason is that when a thread removes a node from a data
structure, without using a lock to exclude other threads, it never knows
if another thread still looking at the node. The algorithms are usually
designed so that the other thread will perform a failing compare-
exchange on a field in the removed node, and thus know to retry.
Unfortunately, if in the meantime the node is handed to free, the field
might be coincidentally set to the value that the compare-exchange
expects to see.
The solution is to use a garbage collector or mini-collector like
hazard pointers. Alternatively you may associate a free list of nodes with
the data structure and not free any nodes until the data structure itself
is freed.

Recommendations
Non-blocking algorithms are currently a hot topic in research. Their big
advantage is avoiding lock pathologies. Their primary disadvantage is that
they are much more complicated than their locked counterparts. Indeed,
the discovery of a lockless algorithm is often worthy of a conference
paper. Non-blocking algorithms are difficult to verify. At least one
incorrect algorithm has made its way into a conference paper. Non-
experts should consider the following advice:
Atomic increment, decrement, and fetch-and-add are generally
safe to use in an intuitive fashion.
The fetch-and-op idiom is generally safe to use with operations
that are commutative and associative.
The creation of non-blocking algorithms for linked data
structures should be left to experts. Use algorithms from the
peer-reviewed literature. Be sure to understand any memory
reclamation issues.
Otherwise, for now, stick with locks. Avoid having more runnable
software threads than hardware threads, and design programs to avoid
lock contention. This way, the problems solved by non-blocking
algorithms will not come up in the first place.
192 Multi-Core Programming

Thread-safe Functions and Libraries

The Foo example in Figure 7.2 underscores the importance of
documenting thread safety. Defining a routine like Foo that updates
unprotected hidden shared state is a poor programming practice. In
general, routines should be thread safe; that is, concurrently callable by
clients. However, complete thread safety is usually unrealistic, because it
would require that every call do some locking, and performance would
be pathetic. Instead, a common convention is to guarantee that instance
routines are thread safe when called concurrently on different objects,
but not thread safe when called concurrently on the same object.
This convention is implicit when objects do not share state. For
objects that do share state, the burden falls on the implementer to
protect the shared state. Figure 7.14 shows a reference-counted
implementation of strings where the issue arises. From the clients
viewpoint, each string object is a separate string object, and thus threads
should be able to concurrently operate on each object. In the underlying
implementation, however, a string object is simply a pointer to a shared
object that has the string data, and a reference count of the number of
string objects that point to it. The implementer should ensure that
concurrent accesses do not corrupt the shared state. For example, the
updates to the reference count should use atomic operations.

Figure 7.14 Implementer Should Ensure Thread Safety of Hidden Shared State

When defining interfaces, care should be taken to ensure that they

can be implemented efficiently in a thread-safe manner. Interfaces should
 Chapter 7: Solutions to Common Parallel Programming Problems 193

not update hidden global state, because with multiple threads, it may not
be clear whose global state is being updated. The C library function
strtok is one such offender. Clients use it to tokenize a string. The first
call sets the state of a hidden parser, and each successive call advances
the parser. The hidden parser state makes the interface thread unsafe.
Thread safety can be obtained by having the implementation put the
parser in thread-local storage. But this introduces the complexity of a
threading package into something that really should not need it in the
first place. A thread-safe redesign of strtok would make the parser
object an explicit argument. Each thread would create its own local
parser object and pass it as an argument. That way, concurrent calls
could proceed blissfully without interference.
Some libraries come in thread-safe and thread-unsafe versions. Be sure
to use the thread-safe version for multi-threaded code. For example, on
Windows, the compiler option /MD is required to dynamically link with
the thread-safe version of the run-time library. For debugging, the
corresponding option is /MDd, which dynamically links with the debug
version of the thread-safe run-time. Read your compiler documentation
carefully about these kinds of options. Because the compilers date back to
the single-core era, the defaults are often for code that is not thread safe.

Memory Issues
When most people perform calculations by hand, they are limited by
how fast they can do the calculations, not how fast they can read and
write. Early microprocessors were similarly constrained. In recent
decades, microprocessors have grown much faster in speed than in
memory. A single microprocessor core can execute hundreds of
operations in the time it takes to read or write a value in main memory.
Programs now are often limited by the memory bottleneck, not
processor speed. Multi-core processors can exacerbate the problem
unless care is taken to conserve memory bandwidth and avoid memory
contention.

Bandwidth
To conserve bandwidth, pack data more tightly, or move it less
frequently between cores. Packing the data tighter is usually
194 Multi-Core Programming

straightforward, and benefits sequential execution as well. For example,

pack Boolean arrays into one Boolean value per bit, not one value per
byte. Use the shortest integer type that can hold values in the required
range. When declaring structures in C/C++, declare fields in order of
descending size. This strategy tends to minimize the extra padding that
the compiler must insert to maintain alignment requirements, as
exemplified in Figure 7.15.

Figure 7.15 Order Fields by Decreasing Size to Reduce Padding

Some compilers also support #pragma pack directives that pack

structures even more tightly, possibly by removing all padding. Such very
tight packing may be counterproductive, however, because it causes
misaligned loads and stores that may be significantly slower than aligned
loads and stores.

Working in the Cache

Moving data less frequently is a more subtle exercise than packing,
because mainstream programming languages do not have explicit
commands to move data between a core and memory. Data movement
arises from the way the cores read and write memory. There are two
categories of interactions to consider: those between cores and memory,
and those between cores.
Data movement between a core and memory also occurs in single-
core processors, so minimizing data movement benefits sequential
programs as well. There exist numerous techniques. For example, a
 Chapter 7: Solutions to Common Parallel Programming Problems 195

technique called cache-oblivious blocking recursively divides a problem

into smaller and smaller subproblems. Eventually the subproblems
become so small that they each fit in cache. The Fastest Fourier
Transform in the West (Frigo 1997) uses this approach and indeed lives
up to its name. Another technique for reducing the cache footprint is to
reorder steps in the code. Sometimes this is as simple as interchanging
loops. Other times it requires more significant restructuring.
The Sieve of Eratosthenes is an elementary programming exercise that
demonstrates such restructuring and its benefits. Figure 7.16 presents the
Sieve of Eratosthenes for enumerating prime numbers up to n. This version
has two nested loops: the outer loop finds primes, and the inner loop,
inside function Strike, strikes out composite numbers. This version is
unfriendly to cache, because the inner loop is over the full length of array
composite, which might be much larger than what fits in cache.

inline long Strike( bool composite[], long i,

long stride, long limit ) {
for( ; i<=limit; i+=stride )
composite[i] = true;
return i;
}
long CacheUnfriendlySieve( long n ) {
long count = 0;
long m = (long)sqrt((double)n);
bool* composite = new bool[n+1];
memset( composite, 0, n );
for( long i=2; i<=m; ++i )
if( !composite[i] ) {
++count;
// Strike walks array of size n here.
Strike( composite, 2*i, i, n );
}
for( long i=m+1; i<=n; ++i )
if( !composite[i] )
++count;
delete[] composite;
return count;
}

Figure 7.16 Cache-Unfriendly Sieve of Eratosthenes

Figure 7.17 shows how the sieve can be restructured to be cache

friendly. Instead of directly representing the conceptual sieve as one big
array, it represents it as a small window into the conceptual sieve. The
196 Multi-Core Programming

window size is approximately n bytes. The restructuring requires that

the original inner loop be stopped when it reaches the end of a window,
and restarted when processing the next window. The array striker
stores the indices of these suspended loops, and has an element for each
prime up to n . The data structures grow much more slowly than n,
6
and so fit in a 10 byte cache even when n approaches values as large as
11 11
10 . Of course, allocating array composite to hold 10 bytes is
impractical on most machines. The later discussion of multi-threading the
sieve describes how to reduce composite to n bytes instead of
n bytes.

long CacheFriendlySieve( long n ) {

long count = 0;
long m = (long)sqrt((double)n);
bool* composite = new bool[n+1];
memset( composite, 0, n );
long* factor = new long[m];
long* striker = new long[m];
long n_factor = 0;
for( long i=2; i<=m; ++i )
if( !composite[i] ) {
++count;
striker[n_factor] = Strike( composite, 2*i, i, m );
factor[n_factor++] = i;
}
// Chops sieve into windows of size sqrt(n)
for( long window=m+1; window<=n; window+=m ) {
long limit = min(window+m-1,n);
for( long k=0; k<n_factor; ++k )
// Strike walks window of size sqrt(n) here.
striker[k] = Strike( composite, striker[k], factor[k],
limit );
for( long i=window; i<=limit; ++i )
if( !composite[i] )
++count;
}
delete[] striker;
delete[] factor;
delete[] composite;
return count;
}

Figure 7.17 Cache-Friendly Sieve of Eratosthenes

The restructuring introduces extra complexity and bookkeeping

operations. But because processor speed so greatly outstrips memory
 Chapter 7: Solutions to Common Parallel Programming Problems 197

speed, the extra bookkeeping pays off dramatically. Figure 7.18 shows
this performance difference. On this log plot, the cache friendly code has
a fairly straight performance plot, while the cache unfriendly versions
running time steps up from one straight line to another when n reaches
6
approximately 10 . The step is characteristic of algorithms that transition
from running in cache to running out of cache as the problem size
increases. The restructured version is five times faster than the original
version when n significantly exceeds the cache size, despite the extra
processor operations required by the restructuring.

Figure 7.18 Performance Difference between Figure 7.16 and Figure 7.17

Memory Contention
For multi-core programs, working within the cache becomes trickier,
because data is not only transferred between a core and memory, but
also between cores. As with transfers to and from memory, mainstream
programming languages do not make these transfers explicit. The
transfers arise implicitly from patterns of reads and writes by different
cores. The patterns correspond to two types of data dependencies:
Read-write dependency. A core writes a cache line, and then a
different core reads it.
Write-write dependency. A core writes a cache line, and then a
different core writes it.
An interaction that does not cause data movement is two cores
repeatedly reading a cache line that is not being written. Thus if multiple
cores only read a cache line and do not write it, then no memory
198 Multi-Core Programming

bandwidth is consumed. Each core simply keeps its own copy of the
cache line.
To minimize memory bus traffic, minimize core interactions by
minimizing shared locations. Hence, the same patterns that tend to
reduce lock contention also tend to reduce memory traffic, because it is
the shared state that requires locks and generates contention. Letting
each thread work on its own local copy of the data and merging the data
after all threads are done can be a very effective strategy.
Consider writing a multi-threaded version of the function
CacheFriendlySieve from Figure 7.17. A good decomposition for this
problem is to fill the array factor sequentially, and then operate on the
windows in parallel. The sequential portion takes time O( n ), and
hence has minor impact on speedup for large n. Operating on the
windows in parallel requires sharing some data. Looking at the nature of
the sharing will guide you on how to write the parallel version.
The array factor is read-only once it is filled. Thus each thread
can share the array.
The array composite is updated as primes are found. However,
the updates are made to separate windows, so they are unlikely
to interfere except at window boundaries that fall inside a cache
line. Better yet, observe that the values in the window are used
only while the window is being processed. The array composite
no longer needs to be shared, and instead each thread can have a
private portion that holds only the window of interest. This
change benefits the sequential version too, because now the
space requirements for the sieve have been reduced from O(n) to
11
O( n ). The reduction in space makes counting primes up to 10
possible on even a 32-bit machine.
The variable count is updated as primes are found. An atomic
increment could be used, but that would introduce memory
contention. A better solution, as shown in the example, is to give
each thread perform a private partial count, and sum the partial
counts at the end.
The array striker is updated as the window is processed. Each
thread will need its own private copy. The tricky part is that
striker induces a loop-carried dependence between windows.
For each window, the initial value of striker is the last value it
had for the previous window. To break this dependence, the
initial values in striker have to be computed from scratch. This
 Chapter 7: Solutions to Common Parallel Programming Problems 199

computation is not difficult. The purpose of striker[k] is to

keep track of the current multiple of factor[k].
The variable base is new in the parallel version. It keeps track of
the start of the window for which striker is valid. If the value
of base differs from the start of the window being processed, it
indicates that the thread must recompute striker from scratch.
The recomputation sets the initial value of striker[k] to the
lowest multiple of factor[k] that is inside or after the window.
Figure 7.19 shows the multi-threaded sieve. A further refinement that
cuts the work in half would be to look for only odd primes. The
refinement was omitted from the examples because it obfuscates
understanding of the multi-threading issues.

long ParallelSieve( long n ) {

long count = 0;
long m = (long)sqrt((double)n);
long n_factor = 0;
long* factor = new long[m];

#pragma omp parallel

{
bool* composite = new bool[m+1];
long* striker = new long[m];

#pragma omp single

{
memset( composite, 0, m );
for( long i=2; i<=m; ++i )
if( !composite[i] ) {
++count;
Strike( composite, 2*i, i, m );
factor[n_factor++] = i;
}
}
long base = -1;

#pragma omp for reduction (+:count)

for( long window=m+1; window<=n; window+=m ) {
memset( composite, 0, m );
if( base!=window ) {
// Must compute striker from scratch.
base = window;
200 Multi-Core Programming

for( long k=0; k<n_factor; ++k )

striker[k] = (base+factor[k]-1)/factor[k] *
factor[k] - base;
}
long limit = min(window+m-1,n) - base;
for( long k=0; k<n_factor; ++k )
striker[k] = Strike( composite, striker[k],
factor[k], limit ) - m;
for( long i=0; i<=limit; ++i )
if( !composite[i] )
++count;
base += m;
}
delete[] striker;
delete[] composite;
}
delete[] factor;
return count;
}

Figure 7.19 Parallel Sieve of Eratosthenes

Cache-related Issues
As remarked earlier in the discussion of time-slicing issues, good
performance depends on processors fetching most of their data from
cache instead of main memory. For sequential programs, modern caches
generally work well without too much thought, though a little tuning
helps. In parallel programming, caches open up some much more serious
pitfalls.

False Sharing
The smallest unit of memory that two processors interchange is a cache
line or cache sector. Two separate caches can share a cache line when
they both need to read it, but if the line is written in one cache, and read
in another, it must be shipped between caches, even if the locations of
interest are disjoint. Like two people writing in different parts of a log
book, the writes are independent, but unless the book can be ripped
apart, the writers must pass the book back and forth. In the same way,
two hardware threads writing to different locations contend for a cache
sector to the point where it becomes a ping-pong game.
 Chapter 7: Solutions to Common Parallel Programming Problems 201

Figure 7.20 illustrates such a ping-pong game. There are two threads,
each running on a different core. Each thread increments a different
location belonging to the same cache line. But because the locations
belong to the same cache line, the cores must pass the sector back and
forth across the memory bus.

Figure 7.20 Cache Line Ping Ponging Caused by False Sharing

Figure 7.21 shows how bad the impact can be for a generalization of
Figure 7.20. Four single-core processors, each enabled with Hyper-
Threading Technology (HT Technology), are used to give the flavor of a
hypothetical future eight-core system. Each hardware thread increments
a separate memory location. The ith thread repeatedly increments
x[i*stride]. The performance is worse when the locations are
adjacent, and improves as they spread out, because the spreading puts
the locations into more distinct cache lines. Performance improves
sharply at a stride of 16. This is because the array elements are 4-byte
integers. The stride of 16 puts the locations 16 4 = 64 bytes apart. The
data is for a Pentium 4 based processor with a cache sector size of
64 bytes. Hence when the locations were 64 bytes part, each thread is
202 Multi-Core Programming

hitting on a separate cache sector, and the locations become private to

each thread. The resulting performance is nearly one hundredfold better
than when all threads share the same cache line.

Figure 7.21 Performance Impact of False Sharing

Avoiding false sharing may require aligning variables or objects in

memory on cache line boundaries. There are a variety of ways to force
alignment. Some compilers support alignment pragmas. The Windows
compilers have a directive __declspec(align(n)) that can be used to
specify n-byte alignment. Dynamic allocation can be aligned by allocating
extra pad memory, and then returning a pointer to the next cache line in
the block. Figure 7.22 shows an example allocator that does this. Function
CacheAlignedMalloc uses the word just before the aligned block to store a
pointer to the true base of the block, so that function CacheAlignedFree
can free the true block. Notice that if malloc returns an aligned pointer,
CacheAlignedMalloc still rounds up to the next cache line, because it
needs the first cache line to store the pointer to the true base.
 Chapter 7: Solutions to Common Parallel Programming Problems 203

It may not be obvious that there is always enough room before the
aligned block to store the pointer. Sufficient room depends upon two
assumptions:
A cache line is at least as big as a pointer.
A malloc request for at least a cache lines worth of bytes returns
a pointer aligned on boundary that is a multiple of
sizeof(char*).

These two conditions hold for IA-32 and Itanium-based systems. Indeed,
they hold for most architecture because of alignment restrictions
specified for malloc by the C standard.

// Allocate block of memory that starts on cache line

void* CacheAlignedMalloc( size_t bytes, void* hint ) {
size_t m = (cache line size in bytes);
assert( (m & m-1)==0 ); // m must be power of 2
char* base = (char*)malloc(m+bytes);

// Round pointer up to next line

char * result = (char*)((UIntPtr)(base+m)&-m);

// Record where block actually starts.

((char**)result)[-1] = base;

return result;
}

// Free block allocated by CacheAlignedMalloc

void CacheAlignedFree( void* p ) {

// Recover where block actually starts

char* base = ((byte**)p)[-1];

// Failure of following assertion indicates memory

// was not allocated with CacheAlignedMalloc.
assert( (void*)((UIntPtr)
(base+NFS_LineSize)&-NFS_LineSize) == p);
free( base );
}

Figure 7.22 Memory Allocator that Allocates Blocks Aligned on Cache Line
Boundaries

The topic of false sharing exposes a fundamental tension between

efficient use of a single-core processor and efficient use of a multi-core
204 Multi-Core Programming

processor. The general rule for efficient execution on a single core is to

pack data tightly, so that it has as small a footprint as possible. But on a
multi-core processor, packing shared data can lead to a severe penalty
from false sharing. Generally, the solution is to pack data tightly, give
each thread its own private copy to work on, and merge results
afterwards. This strategy extends naturally to task stealing. When a thread
steals a task, it can clone the shared data structures that might cause
cache line ping ponging, and merge the results later.

Memory Consistency
At any given instant in time in a sequential program, memory has a well
defined state. This is called sequential consistency. In parallel programs,
it all depends upon the viewpoint. Two writes to memory by a hardware
thread may be seen in a different order by another thread. The reason is
that when a hardware thread writes to memory, the written data goes
through a path of buffers and caches before reaching main memory.
Along this path, a later write may reach main memory sooner than an
earlier write. Similar effects apply to reads. If one read requires a fetch
from main memory and a later read hits in cache, the processor may
allow the faster read to pass the slower read. Likewise, reads and writes
might pass each other. Of course, a processor has to see its own reads
and writes in the order it issues them, otherwise programs would break.
But the processor does not have to guarantee that other processors see
those reads and writes in the original order. Systems that allow this
reordering are said to exhibit relaxed consistency.
Because relaxed consistency relates to how hardware threads observe
each others actions, it is not an issue for programs running time sliced
on a single hardware thread. Inattention to consistency issues can result
in concurrent programs that run correctly on single-threaded hardware,
or even hardware running with HT Technology, but fail when run on
multi-threaded hardware with disjoint caches.
The hardware is not the only cause of relaxed consistency. Compilers
are often free to reorder instructions. The reordering is critical to most
major compiler optimizations. For instance, compilers typically hoist
loop-invariant reads out of a loop, so that the read is done once per loop
instead of once per loop iteration. Language rules typically grant the
compiler license to presume the code is single-threaded, even if it is not.
This is particularly true for older languages such as Fortran, C, and
C++ that evolved when parallel processors were esoteric. For recent
languages, such as Java and C#, compilers must be more circumspect,
 Chapter 7: Solutions to Common Parallel Programming Problems 205

but only when the keyword volatile is present. Unlike hardware

reordering, compiler reordering can affect code even when it is running
time sliced on a single hardware thread. Thus the programmer must be
on the lookout for reordering by the hardware or the compiler.

Current IA-32 Architecture

IA-32 approximates sequential consistency, because it evolved in the
single-core age. The virtue is how IA-32 preserves legacy software.
Extreme departures from sequential consistency would have broken old
code. However, adhering to sequential consistency would have yielded
poor performance, so a balance had to be struck. For the most part, the
balance yields few surprises, yet achieves most of the possible
performance improvements (Hill 1998). Two rules cover typical
programming:
Relaxation for performance. A thread sees other threads reads
and writes in the original order, except that a read may pass a
write to a different location. This reordering rule allows a thread
to read from its own cache even if the read follows a write to
main memory. This rule does not cover nontemporal writes,
which are discussed later.
Strictness for correctness. An instruction with the LOCK prefix
acts as a memory fence. No read or write may cross the fence.
This rule stops relaxations from breaking typical synchronization
idioms based on the LOCK instructions. Furthermore, the
instruction xchg has an implicit LOCK prefix in order to preserve
old code written before the LOCK prefix was introduced.
This slightly relaxed memory consistency is called processor order. For
efficiency, the IA-32 architecture also allows loads to pass loads but hides
this from the programmer. But if the processor detects that the
reordering might have a visible effect, it squashes the affected
instructions and reruns them. Thus the only visible relaxation is that
reads can pass writes.
The IA-32 rules preserve most idioms, but ironically break the
1
textbook algorithm for mutual exclusion called Dekkers Algorithm .
This algorithm enables mutual exclusion for processors without special
atomic instructions. Figure 7.23(a) demonstrates the key sequence in
Dekkers Algorithm. Two variables X and Y are initially zero. Thread 1

1
The first published software-only, two-process mutual exclusion algorithm.
206 Multi-Core Programming

writes X and reads Y. Thread 2 writes Y and reads X. On a sequentially

consistent machine, no matter how the reads and writes are
interleaved, no more than one of the threads reads a zero. The thread
reading the zero is the one allowed into the exclusion region. On IA-32,
and just about every other modern processor, both threads might read
0, because the reads might pass the writes. The code behaves as if
written in Figure 7.23(b).

Figure 7.23 Why Relaxed Consistency Breaks Dekkers Algorithm

Figure 7.23(c) shows how make the sequence work by inserting

explicit memory fence instructions. The fences keep the reads from
passing the writes. Table 7.1 summarizes the three types of IA-32 fence
instructions.

Table 7.1 Types of IA-32 Fence Instructions

Mnemonic Name Description
mfence Memory fence neither reads nor writes may cross
lfence load fence reads may not cross
sfence Store fence writes may not cross
 Chapter 7: Solutions to Common Parallel Programming Problems 207

The fences serve to tighten memory ordering when necessary for

correctness. The order of writes can be loosened with nontemporal store
instructions, which are not necessarily seen by other processors in the
order they were issued by the issuing processor. Some IA-32 string
operations, such as MOVS and STOS, can be nontemporal. The looser
memory ordering allows the processor to maximize bus efficiency by
combining writes. However, the processor consuming such data might
not be expecting to see the writes out of order, so the producer should
issue a sfence before signaling the consumer that the data is ready.
IA-32 also allows memory consistency rules to be varied for specific
memory ranges. For instance, a range with write combining permits
the processor to temporarily record writes in a buffer, and commit the
results to cache or main memory later in a different order. Such a range
behaves as if all stores are nontemporal. In practice, in order to
preserve legacy code, most environments configure IA-32 systems to
use processor order, so the page-by-page rules apply only in special

environments. Section 7.2 of Volume 2 of IA-32 Intel Architecture
Software Developers Manual describes the memory ordering rules in
more detail.

Itanium Architecture
The Itanium architecture had no legacy software to preserve, and thus
could afford a cutting-edge relaxed memory model. The model
theoretically delivers higher performance than sequential consistency by
giving the memory system more freedom of choice. As long as locks are
properly used to avoid race conditions, there are no surprises. However,
programmers writing multiprocessor code with deliberate race
conditions must understand the rules. Though far more relaxed than
IA-32, the rules for memory consistency on Itanium processors are
simpler to remember because they apply uniformly. Furthermore,
compilers for Itanium-based systems interpret volatile in a way that
makes most idioms work.
Figure 7.24(a) shows a simple and practical example where the rules
come into play. It shows two threads trying to pass a message via
memory. Thread 1 writes a message into variable Message, and Thread 2
reads the message. Synchronization is accomplished via the flag
IsReady. The writer sets IsReady after it writes the message. The reader
208 Multi-Core Programming

busy waits for IsReady to be set, and then reads the message. If the
writes or reads are reordered, then Thread 2 may read the message
before Thread 1 is done writing it. Figure 7.24(b) shows how the Itanium
architecture may reorder the reads and writes. The solution is to declare
the flag IsReady as volatile, as shown in 7.24(c). Volatile writes
are compiled as store with release and volatile reads are compiled as
load with acquire. Memory operations are never allowed to move
downwards over a release or upwards over an acquire, thus
enforcing the necessary orderings.

Figure 7.24 Use of Volatile Keyword for Itanium Architecture

The details of the Itanium architectures relaxed memory model

can be daunting, but the two of the idioms over most practice. Figure
7.25 illustrates these two idioms. The animals represent memory
operations whose movement is constrained by animal trainers who
represent acquire and release fences. The first idiom is message
passing, which is a generalization of Figure 7.24. A sender thread
 Chapter 7: Solutions to Common Parallel Programming Problems 209

writes some data, and then signals a receiver thread that it is ready by
modifying a flag location. The modification might be a write, or some
other atomic operation. As long as the sender performs a release
operation after writing the data, and the receiver performs an acquire
operation before reading the data, the desired ordering will be
maintained. Typically, these conditions are guaranteed by declaring
the flag volatile, or using an atomic operation with the desired
acquire/release characteristics.

Operations by Sending Thread Operations by Receiving Thread

Release Acquire

Write Operations for Message Read Operations for Message

Time
Message Passing Idiom

Cage Boundary
Memory Operations

Acquire Release

Time
Cage Idiom

Figure 7.25 Two Common Idioms for Using Shared Memory without a Lock

The second idiom is memory cage. A memory cage starts with an

acquire fence and ends in a release fence. These fences keep any memory
operations inside the cage from escaping. However, be aware that
memory cages keep things inside from getting out, and not vice-versa. It
is possible for disjoint cages to become overlapped by instruction
reordering, because an acquire that begins a cage can float backwards
over the release that ends a previous cage. For similar reasons, trying to
fix Dekkers Algorithm with acquiring reads and releasing writes does not
fix the algorithmthe fix needs to stop reads from floating backwards
over writes, but acquiring reads can nonetheless float backwards over
releasing writes. The proper fix is to add a full memory fence, for
instance, call the __memory_barrier() intrinsic.
210 Multi-Core Programming

A subtle example of fencing is the widely used double-check idiom.

The idiom is commonly used for lazy initialization in multi-threaded code.
Figure 7.26 shows a correct implementation of double check for the
Itanium architecture. The critical feature is declaring the flag as volatile
so that the compiler will insert the correct acquire and release fences.
Double-check is really the message-passing idiom, where the message is
the initialized data structure. This implementation is not guaranteed to be
correct by the ISO C and C++ standards, but is nonetheless correct for
the Itanium architecture because the Itanium processors interpretation
of volatile reads and writes implies fences.

Figure 7.26 Use of Volatile in Double Check Idiom on Itanium Architecture

A common analysis error is to think that the acquire fence between

the outer if and read data structure is redundant, because it would
seem that the hardware must perform the if before the read data
structure. But an ordinary read could in fact be hoisted above the if
were it on the same cache line as another read before the if. Likewise,
without the fence, an aggressive compiler might move the read upwards
over the if as a speculative read. The acquire fence is thus critical.

High-level Languages
When writing portable code in a high-level language, the easiest way to
deal with memory consistency is through the languages existing
synchronization primitives, which normally have the right kind of fences
built in. Memory consistency issues appear only when programmers roll
 Chapter 7: Solutions to Common Parallel Programming Problems 211

their own synchronization primitives. If you must roll your own

synchronization, the rules depend on the language and hardware. Here
are some guidelines:
C and C++. There is no portable solution at present. The ISO C++
committee is considering changes that would address the issue.
For Windows compilers for IA-32, use inline assembly code to
embed fence instructions. For the Itanium processor family, try to
stick to the message passing and cage idioms, and declare the
appropriate variables as volatile.
.NET. Use volatile declarations as for the Itanium architecture and
the code should be portable to any architecture.
Java. The recent JSR-133 revision of the Java memory makes it
similar to Itanium architecture with .NET, so likewise, use volatile
declarations.

Avoiding Pipeline Stalls on IA-32

When writing a parallel program for performance, first get the
decomposition right. Then tune for cache usage, including avoidance of
false sharing. Then, as a last measure, if trying to squeeze the last cycles
out, concern yourself with the processors pipeline. The Pentium 4 and
Pentium D processors have deep pipelines that permit high execution
rates of instructions typical of single-threaded code. The execution units
furthermore reorder instructions so that instructions waiting on memory
accesses do not block other instructions. Deep pipelines and out of order
execution are usually a good thing, but make some operations relatively
expensive.
Particularly expensive are serializing instructions. These are
instructions that force all prior instructions to complete before any
subsequent instructions. Common serializing instructions include those
with the LOCK prefix, memory fences, and the CPUID instruction. The
XCHG instruction on memory is likewise serializing, even without the
LOCK prefix. These instructions are essential when serving their
purpose, but it can pay to avoid them, or at least minimize them, when
such alternatives exist.
On processors with HT Technology, spin waits can be a problem
because the spinning thread might consume all the hardware resources.
In the worst case, it might starve the thread on which the spinner is
212 Multi-Core Programming

waiting! On the Pentium 4 processor and later processors, the solution is

to insert a PAUSE instruction. On Itanium processors, the similar
instruction is HINT 0. These instructions notify the hardware that the
thread is waiting; that is, that hardware resources should be devoted to
other threads. Furthermore, on IA-32, spinning on a read can consume
bus bandwidth, so it is typically best to incorporate exponential backoff
too. Figure 7.27 shows a spin-wait with a PAUSE instruction
incorporated. In more complicated waits based on exponential backoff,
the PAUSE instruction should go in the delay loop.

while( !IsReady )
_asm pause;

R2 = Message;

Figure 7.27 Busy-wait Loop with IA-32 Pause Instruction

Data Organization for High Performance

The interactions of memory, cache, and pipeline can be daunting. It may
help to think of a programs locations as divided into four kinds of
locality:
Thread private. These locations are private to a given thread and
never shared with other threads. A hardware thread tends to
keep this memory in cache, as long as it fits. Hence accesses to
thread private locations tend to be very fast and not consume bus
bandwidth.
Thread shared read only. These locations are shared by multiple
threads, but never written by those threads. Lookup tables are a
common example. Because there are no writes, a hardware
thread tends to keep its own copy in cache, as long as it fits.
Exclusive access. These locations are read and written, but
protected by a lock. Once a thread acquires the lock and starts
operating on the data, the locations will migrate into cache. Once
a thread releases the lock, the locations will migrate back to
memory or to the next hardware thread that acquires the lock.
 Chapter 7: Solutions to Common Parallel Programming Problems 213

Wild West. These locations are read and written by

unsynchronized threads. Depending upon the lock implementa-
tion, these locations may include the lock objects themselves,
because by their nature, locks are accessed by unsynchronized
threads that the lock will synchronize. Whether a lock object
counts as part of the Wild West depends upon whether the lock
object holds the real guts of the lock, or is just a pointer off to
the real guts.
A locations locality may change as the program runs. For instance, a
thread may create a lookup table privately, and then publish its location
to other threads so that it becomes a read-only table.
A good decomposition favors thread-private storage and thread-shared
read-only storage, because these have low impact on the bus and do not
need synchronization. Furthermore, locations of a given locality should not
be mixed on the same cache line, because false sharing issues arise. For
example, putting thread-private data and Wild West data on the same line
hinders access to the thread-private data as the Wild West accesses ping
pong the line around. Furthermore, Wild West locations are often
candidates for putting on a separate cache line, unless the locations tend to
be accessed by the same thread at nearly the same time, in which case
packing them onto the same cache line may help reduce memory traffic.

Key Points
The key to successful parallel programming is choosing a good program
decomposition. Keep the following points in mind when choosing a
decomposition:
Match the number of runnable software threads to the available
hardware threads. Never hard-code the number of threads into
your program; leave it as a tuning parameter.
Parallel programming for performance is about finding the zone
between too little and too much synchronization. Too little
synchronization leads to incorrect answers. Too much
synchronization leads to slow answers.
Use tools like Intel Thread Checker to detect race conditions.
Keep locks private. Do not hold a lock while calling another
packages code.
Avoid deadlock by acquiring locks in a consistent order.
214 Multi-Core Programming

Consider memory bandwidth and contention issues when picking

a decomposition for parallel execution. Pack data tightly to
minimize bandwidth and cache footprint. Put data meant for
different processors on different cache lines. Separate shared
read-only data from writable data.
Spread out lock contention by using multiple distributed locks
where possible.
Lockless algorithms have both advantages and disadvantages.
They are particularly tricky in languages without garbage
collection. Tread with caution.
Cache lines are the quanta of information interchange between
hardware threads.
If writing your own synchronization code, understand the
memory consistency model for the target platform.
Serializing instructions, such as atomic operation and memory
fences, are useful when needed, but relatively expensive
compared to other instructions.
 Chapter 8
Multi-threaded
Debugging
Techniques
D ebugging multi-threaded applications can be a challenging task. The
increased complexity of multi-threaded programs results in a large
number of possible states that the program may be in at any given time.
Determining the state of the program at the time of failure can be
difficult; understanding why a particular state is troublesome can be even
more difficult. Multi-threaded programs often fail in unexpected ways,
and often in a nondeterministic fashion. Bugs may manifest themselves in
a sporadic fashion, frustrating developers who are accustomed to
troubleshooting issues that are consistently reproducible and predictable.
Finally, multi-threaded applications can fail in a drastic fashion
deadlocks cause an application or worse yet, the entire system, to hang.
Users tend to find these types of failures to be unacceptable.
This chapter examines general purpose techniques that are useful
when debugging multi-threaded applications. Intel has developed a

number of tools, including the Intel Thread Checker, the Intel Thread
Profiler, and the Intel Debugger that help debug and profile multi-
threaded applications. These tools are discussed in Chapter 11.

General Debug Techniques

Regardless of which library or platform that you are developing on,
several general principles can be applied to debugging multi-threaded
software applications.

215
216 Multi-Core Programming

Designing with Debugging in Mind

The first technique for eliminating bugs in multi-threaded code is to avoid
introducing the bug in the first place. Many software defects can be
1
prevented by using proper software development practices. The later a
problem is found in the product development lifecycle, the more
expensive it is to fix. Given the complexity of multi-threaded programs, it
is critical that multi-threaded applications are properly designed up front.
How often have you, as a software developer, experienced the
following situation? Someone on the team that youre working on gets a
great idea for a new product or feature. A quick prototype that illustrates
the idea is implemented and a quick demo, using a trivial use-case, is
presented to management. Management loves the idea and immediately
informs sales and marketing of the new product or feature. Marketing
then informs the customer of the feature, and in order to make a sale,
promises the customer the feature in the next release. Meanwhile, the
engineering team, whose original intent of presenting the idea was to get
resources to properly implement the product or feature sometime in the
future, is now faced with the task of delivering on a customer
commitment immediately. As a result of time constraints, it is often the
case that the only option is to take the prototype, and try to turn it into
production code.
While this example illustrates a case where marketing and
management may be to blame for the lack of following an appropriate
process, software developers are often at fault in this regard as well. For
many developers, writing software is the most interesting part of the
job. Theres a sense of instant gratification when you finish writing your
application and press the run button. The results of all the effort and
hard work appear instantly. In addition, modern debuggers provide a
wide range of tools that allow developers to quickly identify and fix
simple bugs. As a result, many programmers fall into the trap of coding
now, deferring design and testing work to a later time. Taking this
approach on a multi-threaded application is a recipe for disaster for
several reasons:
Multi-threaded applications are inherently more complicated
than single-threaded applications. Hacking out a reliable,
scalable implementation of a multi-threaded application is hard;

1
There are a number of different software development methodologies that are applicable to parallel
programming. For example, parallel programming can be done using traditional or rapid prototyping
(Extreme Programming) techniques.
 Chapter 8: Multi-threaded Debugging Techniques 217

even for experienced parallel programmers. The primary reason

for this is the large number of corner cases that can occur and the
wide range of possible paths of the application. Another
consideration is the type of run-time environment the application
is running on. The access patterns may vary wildly depending on
whether or not the application is running on a single-core or
multi-core platform, and whether or not the platform supports
simultaneous multithreading hardware. These different run-time
scenarios need to be thoroughly thought out and handled to
guarantee reliability in a wide range of environments and use
cases.
Multi-threaded bugs may not surface when running under the
debugger. Multi-threading bugs are very sensitive to the timing of
events in an application. Running the application under the
debugger changes the timing, and as a result, may mask
problems. When your application fails in a test or worse, the
customer environment, but runs reliably under the debugger, it is
almost certainly a timing issue in the code.
While following a software process can feel like a nuisance at times,
taking the wrong approach and not following any process at all is a
perilous path when writing all but the most trivial applications. This
holds true for parallel programs.
While designing your multi-threaded applications, you should keep
these points in mind.
Design the application so that it can run sequentially. An
application should always have a valid means of sequential
execution. The application should be validated in this run mode
first. This allows developers to eliminate bugs in the code that are
not related to threading. If a problem is still present in this mode
of execution, then the task of debugging reduces to single-
threaded debugging.
In many circumstances, it is very easy to generate a sequential
version of an application. For example, an OpenMP application
compiled with one of the Intel compilers can use the openmp-
stubs option to tell the compiler to generate sequential OpenMP
code.
Use established parallel programming patterns. The best
defense against defects is to use parallel patterns that are known
to be safe. Established patterns solve many of the common
218 Multi-Core Programming

parallel programming problems in a robust manner. Reinventing

the wheel is not necessary in many cases.
Include built-in debug support in the application. When trying
to root cause an application fault, it is often useful for
programmers to be able to examine the state of the system at any
arbitrary point in time. Consider adding functions that display the
state of a threador all active threads. Trace buffers, described in
the next section, may be used to record the sequence of accesses
to a shared resource. Many modern debuggers support the
capability of calling a function while stopped at a breakpoint.
This mechanism allows developers to extend the capabilities of
the debugger to suit their particular applications needs.

Code Reviews
Many software processes suggest frequent code reviews as a means
of improving software quality. The complexity of parallel
programming makes this task challenging. While not a replacement
for using well established parallel programming design patterns,
code reviews may, in many cases, help catch bugs in the early stages
of development.
One technique for these types of code reviews is to have individual
reviewers examine the code from the perspective of one of the threads
in the system. During the review, each reviewer steps through the
sequence of events as the actual thread would. Have objects that
represent the shared resources of the system available and have the
individual reviewers (threads) take and release these resources. This
technique will help you visualize the interaction between different
threads in your system and hopefully help you find bugs before they
manifest themselves in code.
As a developer, when you get the urge to immediately jump into
coding and disregard any preplanning or preparation, you should
consider the following scenarios and ask yourself which situation youd
rather be in. Would you rather spend a few weeks of work up front to
validate and verify the design and architecture of your application, or
would you rather deal with having to redesign your product when you
find it doesnt scale? Would you rather hold code reviews during
development or deal with the stress of trying to solve mysterious,
unpredictable showstopper bugs a week before your scheduled ship
date? Good software engineering practices are the key to writing reliable
 Chapter 8: Multi-threaded Debugging Techniques 219

software applications. Nothing is new, mysterious, or magical about

writing multi-threaded applications. The complexity of this class of
applications means that developers must be conscious of these
fundamental software engineering principles and be diligent in following
them.

Extending your ApplicationUsing Trace Buffers

Chapter 7 identified two categories of bugs found in multi-threaded
applications: synchronization bugs and performance bugs.
Synchronization bugs include race conditions and deadlocks that
cause unexpected and incorrect behavior. Performance bugs arise
from unnecessary thread overhead due to thread creation or context
switch overhead, and memory access patterns that are suboptimal
for a given processors memory hierarchy. The application returns
the correct results, but often takes too long to be usable. This
chapter focuses on debugging synchronization bugs that cause
applications to fail.
In order to find the cause of these types of bugs, two pieces of
information are needed:
1. Which threads are accessing the shared resource at the time of
the failure.
2. When the access to the shared resource took place.
In many cases, finding and fixing synchronization bugs involves code
inspection. A log or trace of the different threads in the application and
the pattern in which they accessed the shared resources of the code
helps narrow down the problematic code sections. One simple data
structure that collects this information is the trace buffer.
A trace buffer is simply a mechanism for logging events that the
developer is interested in monitoring. It uses an atomic counter that
keeps track of the current empty slot in the array of event records. The
type of information that each event can store is largely up to the
developer. A sample implementation of a trace buffer, using the Win32
2
threading APIs, is shown in Listing 8.1.

2
In the interest of making the code more readable, Listing 8.1 uses the time() system call to record
system time. Due to the coarse granularity of this timer, most applications should use a high
performance counter instead to keep track of the time in which events occurred.
220 Multi-Core Programming

1 // Circular 1K Trace buffer

2 #define TRACE_BUFFER_SIZE 1024
3
4 typedef struct traceBufferElement
5 {
6 DWORD threadId;
7 time_t timestamp;
8 const char *msg;
9 } traceBufferElement;
10
11 static LONG m_TraceBufferIdx = -1;
12 static traceBufferElement traceBuffer[TRACE_BUFFER_SIZE];
13
14 void InitializeTraceBuffer()
15 {
16 m_TraceBufferIdx = -1;
17
18 /* initialize all entries to {0, 0, NULL} */
19 memset(traceBuffer, 0,
20 TRACE_BUFFER_SIZE*sizeof(traceBufferElement));
21 }
22
23 void AddEntryToTraceBuffer(const char *msg)
24 {
25 LONG idx = 0;
26
27 // Get the index into the trace buffer that this
28 // thread should use
29 idx = InterlockedIncrement(&m_TraceBufferIdx) %
30 TRACE_BUFFER_SIZE;
31
32 // Enter the data into the Trace Buffer
33 traceBuffer[idx].threadId = GetCurrentThreadId();
34 traceBuffer[idx].timestamp = time(NULL);
35 traceBuffer[idx].msg = msg;
36 }
37
38 void PrintTraceBuffer()
39 {
40 int i;
41 printf("Thread ID Timestamp Msg\n");
42 printf("----------|----------|----------------------
43 "-----------------\n");
44
45 // sort by timestamp before printing
46 SortTraceBufferByTimestamp();
 Chapter 8: Multi-threaded Debugging Techniques 221

47 for (i = 0; i < TRACE_BUFFER_SIZE; i++)

48 {
49 if (traceBuffer[i].timestamp == 0)
50 {
51 break;
52 }
53 printf("0x%8.8x|0x%8.8x| %s\n",
54 traceBuffer[i].threadId,
55 traceBuffer[i].timestamp,
56 traceBuffer[i].msg);
57 }
58 }

Listing 8.1 Sample Implementation of a Trace Buffer

Listing 8.1, creates a trace buffer that can store 1,024 events. It stores
these events in a circular buffer. As youll see shortly, once the circular
buffer is full, your atomic index will wrap around and replace the oldest
event. This simplifies your implementation as it doesnt require
dynamically resizing the trace buffer or storing the data to disk. In some
instances, these operations may be desirable, but in general, a circular
buffer should suffice.
Lines 113 define the data structures used in this implementation.
The event descriptor traceBufferElement is defined in lines 49. It
contains three fields: a field to store the thread ID, a timestamp value that
indicates when the event occurred, and a generic message string that is
associated with the event. This structure could include a number of
additional parameters, including the name of the thread.
The trace buffer in Listing 8.1 defines three operations. The first
method, InitializeTraceBuffer(), initializes the resources used by
the trace buffer. The initialization of the atomic counter occurs on line 16.
The atomic counter is initialized to 1. The initial value of this counter is
1 because adding a new entry in the trace buffer requires us to first
increment (line 29) the atomic counter. The first entry should be stored
in slot 0. Once the trace buffer is initialized, threads may call
AddEntryToTraceBuffer() to update the trace buffers with events as
they occur. PrintTraceBuffer() dumps a listing of all the events that
the trace buffer has logged to the screen. This function is very useful
when combined with a debugger that allows users to execute code at a

breakpoint. Both Microsoft Visual Studio and GDB support this
capability. With a single command, the developer can see a log of all the
222 Multi-Core Programming

recent events being monitored, without having to parse a data structure

using the command line or a watch window.
Note that the implementation of the trace buffer in Listing 8.1 logs
events as they are passed into the buffer. This doesnt necessarily
guarantee that the trace buffer will log events exactly as they occur in
time. To illustrate this point, consider the two threads shown in
Listing 8.2.

unsigned __stdcall Thread1(void *)

{
// ... thread initialization
// write global data
m_global = do_work();
AddEntryToTraceBuffer(msg);
// ... finish thread
}

unsigned __stdcall Thread2(void *)

{
// ... thread initialization
// read global data
Thread_local_data = m_global;
AddEntryToTraceBuffer(msg);
// ... finish thread
}

Listing 8.2 Two Threads Logging Events to a Trace Buffer

By now it should be clear what the problem is. A race condition

exists between the two threads and the access to the trace buffer.
Thread1 may write to the global data value and then start logging that
write event in the trace buffer. Meanwhile, Thread2 may read that same
global value after the write, but log this read event before the write
event. Thus, the data in the buffer may not be an accurate reflection of
the actual sequence of events as they occurred in the system.
One potential solution to this problem is to protect the operation
that you want to log and the subsequent trace buffer access with a
synchronization object. A thread, when logging the event, could request
exclusive access to the trace buffer. Once the thread has completed
logging the event, it would then unlock the trace buffer, allowing other
threads to access the buffer. This is shown in Listing 8.3.
 Chapter 8: Multi-threaded Debugging Techniques 223

// This is NOT RECOMMENDED

unsigned __stdcall Thread1(void *)
{
// ... thread initialization
// write global data

LockTraceBuffer();

m_global = do_work();
AddEntryToTraceBuffer(msg);

UnlockTraceBuffer();

// ... finish thread

}

unsigned __stdcall Thread2(void *)

{
// ... thread initialization
// read global data
LockTraceBuffer();

Thread_local_data = m_global;
AddEntryToTraceBuffer(msg);

UnlockTraceBuffer();
// ... finish thread
}

Listing 8.3 Incorrectly Synchronizing Access to the Trace Buffer

There are a number of drawbacks to this technique. Using a

synchronization primitive to protect access to a trace buffer may actually
mask bugs in the code, defeating the purpose of using the trace buffer for
debug. Assume that the bug the developer is tracking down is related to a
missing lock around the read or write access in the thread. By locking
access to the trace buffer, the developer is protecting a critical section of
code that may be incorrectly unprotected. Generally speaking, when
tracking down a race condition, the programmer should avoid
synchronizing access to the trace buffer. If you synchronize access and
your application works, its a clue that there may be a problem in the
synchronization mechanism between those threads.
The preferred method to overcoming this limitation is to log a message
before and after the event occurs. This is demonstrated in Listing 8.4.
224 Multi-Core Programming

unsigned __stdcall Thread1(void *)

{
// ... thread initialization
// write global data
AddEntryToTraceBuffer(before_msg);
m_global = do_work();
AddEntryToTraceBuffer(after_msg);
// ... finish thread
}

unsigned __stdcall Thread2(void *)

{
// ... thread initialization
// read global data
AddEntryToTraceBuffer(before_msg2);
Thread_local_data = m_global;
AddEntryToTraceBuffer(after_msg2);
// ... finish thread
}

Listing 8.4 Preferred Method of Logging Messages with a Trace buffer

By logging a before and after message, a programmer can determine

whether or not the events occurred as expected. If the before and after
messages between the two threads occur in sequence, then the
developer can safely assume that the event was ordered. If the before and
after messages are interleaved, then the order of events is indeterminate;
the events may have happened in either order.
A trace buffer can be used to gather useful data about the sequence
of operations occurring in a multi-threaded application. For other more
difficult problems, more advanced threading debug tools may be
required. These tools are discussed in Chapter 11.

Debugging Multi-threaded Applications in Windows

Most Windows programmers use Microsoft Visual Studio as their primary
integrated development environment (IDE). As part of the IDE, Microsoft
includes a debugger with multi-threaded debug support. This section
examines the different multi-threaded debug capabilities of Visual Studio,
and then demonstrates how they are used.
 Chapter 8: Multi-threaded Debugging Techniques 225

Threads Window
As part of the debugger, Visual Studio provides a Threads window that
lists all of the current threads in the system. From this window, you can:
Freeze (suspend) or thaw (resume) a thread. This is useful
when you want to observe the behavior of your application
without a certain thread running.
Switch the current active thread. This allows you to manually
perform a context switch and make another thread active in the
application.
Examine thread state. When you double-click an entry in the
Threads window, the source window jumps to the source line
that the thread is currently executing. This tells you the threads
current program counter. You will be able to examine the state of
local variables within the thread.
The Threads window acts as the command center for examining and
controlling the different threads in an application.

Tracepoints
As previously discussed, determining the sequence of events that lead to
a race condition or deadlock situation is critical in determining the root
cause of any multi-thread related bug. In order to facilitate the logging of
events, Microsoft has implemented tracepoints as part of the debugger
for Visual Studio 2005.
Most developers are familiar with the concept of a breakpoint. A
tracepoint is similar to a breakpoint except that instead of stopping
program execution when the applications program counter reaches that
point, the debugger takes some other action. This action can be printing
a message or running a Visual Studio macro.
Enabling tracepoints can be done in one of two ways. To create a
new tracepoint, set the cursor to the source line of code and select
Insert Tracepoint. If you want to convert an existing breakpoint to a
tracepoint, simply select the breakpoint and pick the When Hit option
from the Breakpoint submenu. At this point, the tracepoint dialog
appears.
When a tracepoint is hit, one of two actions is taken based on the
information specified by the user. The simplest action is to print a
message. The programmer may customize the message based on a set of
predefined keywords. These keywords, along with a synopsis of what
226 Multi-Core Programming

gets printed, are shown in Table 8.1. All values are taken at the time the
tracepoint is hit.

Table 8.1 Tracepoint Keywords

Keyword Evaluates to
$ADDRESS The address of the instruction
$CALLER The name of the function that called this function
$CALLSTACK The state of the callstack
$FUNCTION The name of the current function
$PID The ID of the process
$PNAME The name of the process
$TID The ID of the thread
$TNAME The name of the thread

In addition to the predefined values in Table 8.1, tracepoints also give

you the ability to evaluate expressions inside the message. In order to do
this, simply enclose the variable or expression in curly braces. For
example, assume your thread has a local variable threadLocalVar that
youd like to have displayed when a tracepoint is hit. The expression
youd use might look something like this:
Thread: $TNAME local variables value is {threadLocalVar}.

Breakpoint Filters
Breakpoint filters allow developers to trigger breakpoints only when
certain conditions are triggered. Breakpoints may be filtered by machine
name, process, and thread. The list of different breakpoint filters is
shown in Table 8.2.

Table 8.2 Breakpoint Filter Options

Filter Description
MachineName Specifies that the breakpoint should only be triggered on
certain machines
ProcessId Limit breakpoint to process with the matching ID
ProcessName Limit breakpoint to process with matching name
ThreadId Limit breakpoint to thread with matching ID
ThreadName Limit breakpoint to thread with matching name
 Chapter 8: Multi-threaded Debugging Techniques 227

Breakpoint filters can be combined to form compound statements.

Three logic operators are supported: !(NOT), &(AND), and ||(OR).

Naming Threads
When debugging a multi-threaded application, it is often useful to assign
unique names to the threads that are used in the application. In
Chapter 5, you learned that assigning a name to a thread in a managed
application was as simple as setting a property on the thread object. In
this environment, it is highly recommended that you set the name field
when creating the thread, because managed code provides no way to
identify a thread by its ID.
In native Windows code, a thread ID can be directly matched to an
individual thread. Nonetheless, keeping track of different thread IDs
makes the job of debugging more difficult; it can be hard to keep track of
individual thread IDs. An astute reader might have noticed in Chapter 5
the conspicuous absence of any sort of name parameter in the methods
used to create threads. In addition, there was no function provided to get
or set a thread name. It turns out that the standard thread APIs in Win32
lack the ability to associate a name with a thread. As a result, this
association must be made by an external debugging tool.
Microsoft has enabled this capability through predefined exceptions
built into their debugging tools. Applications that want to see a thread
referred to by name need to implement a small function that raises an
exception. The exception is caught by the debugger, which then takes
the specified name and assigns it to the associated ID. Once the
exception handler completes, the debugger will use the user-supplied
name from then on.
The implementation of this function can be found on the Microsoft
Developer Network (MSDN) Web site at msdn.microsoft.com by
searching for: setting a thread name (unmanaged). The function,
named SetThreadName(), takes two arguments. The first argument is
the thread ID. The recommended way of specifying the thread ID is to
send the value -1, indicating that the ID of the calling thread should be
used. The second parameter is the name of the thread. The
SetThreadName() function calls RaiseException(), passing in a
special thread exception code and a structure that includes the thread
ID and name parameters specified by the programmer.
Once the application has the SetThreadName() function defined,
the developer may call the function to name a thread. This is shown in
228 Multi-Core Programming

3
Listing 8.5. The function Thread1 is given the name Producer,
indicating that it is producing data for a consumer. Note that the function
is called at the start of the thread, and that the thread ID is specified as -1.
This indicates to the debugger that it should associate the calling thread
with the associated ID.

unsigned __stdcall Thread1(void *)

{
int i, x = 0; // arbitrary local variable declarations
SetThreadName(-1, "Producer");

// Thread logic follows

}

Listing 8.5 Using SetThreadName to Name a Thread

Naming a thread in this fashion has a couple of limitations. This

technique is a debugger construct; the OS is not in any way aware of the
name of the thread. Therefore, the thread name is not available to anyone
other than the debugger. You cannot programmatically query a thread
for its name using this mechanism. Assigning a name to a thread using
this technique requires a debugger that supports exception number
0x406D1388. Both Microsofts Visual Studio and WinDbg debuggers
support this exception. Despite these limitations, it is generally advisable
to use this technique where supported as it makes using the debugger
and tracking down multi-threaded bugs much easier.

Putting It All Together

Lets stop for a minute and take a look at applying the previously
discussed principles to a simplified real-world example. Assume that you
are writing a data acquisition application. Your design calls for a
producer thread that samples data from a device every second and stores
the reading in a global variable for subsequent processing. A consumer
thread periodically runs and processes the data from the producer. In
order to prevent data corruption, the global variable shared by the
producer and consumer is protected with a Critical Section. An example
of a simple implementation of the producer and consumer threads is
shown in Listing 8.6. Note that error handling is omitted for readability.

3
Admittedly the function name Thread1 should be renamed to Producer as well, but is left
somewhat ambiguous for illustration purposes.
 Chapter 8: Multi-threaded Debugging Techniques 229

1 static int m_global = 0;

2 static CRITICAL_SECTION hLock; // protect m_global
3
4 // Simple simulation of data acquisition
5 void sample_data()
6 {
7 EnterCriticalSection(&hLock);
8 m_global = rand();
9 LeaveCriticalSection(&hLock);
10 }
11
12 // This function is an example
13 // of what can be done to data
14 // after collection
15 // In this case, you update the display
16 // in real time
17 void process_data()
18 {
19 EnterCriticalSection(&hLock);
20 printf("m_global = 0x%x\n", m_global);
21 LeaveCriticalSection(&hLock);
22 }
23
24 // Producer thread to simulate real time
25 // data acquisition. Collect 30 s
26 // worth of data
27 unsigned __stdcall Thread1(void *)
28 {
29 int count = 0;
30 SetThreadName(-1, "Producer");
31 while (1)
32 {
33 // update the data
34 sample_data();
35
36 Sleep(1000);
37 count++;
38 if (count > 30)
39 break;
40 }
41 return 0;
42 }
43
44 // Consumer thread
45 // Collect data when scheduled and
46 // process it. Read 30 s worth of data
47 unsigned __stdcall Thread2(void *)
48 {
49 int count = 0;
230 Multi-Core Programming

50 SetThreadName(-1, "Consumer");
51 while (1)
52 {
53 process_data();
54
55 Sleep(1000);
56 count++;
57 if (count > 30)
58 break;
59 }
60 return 0;
61 }

Listing 8.6 Simple Data Acquisition Device

The producer samples data on line 34 and the consumer processes

the data in line 53. Given this relatively simple situation, it is easy to
verify that the program is correct and free of race conditions and
deadlocks. Now assume that the programmer wants to take advantage of
an error detection mechanism on the data acquisition device that
indicates to the user that the data sample collected has a problem. The
changes made to the producer thread by the programmer are shown in
Listing 8.7.

void sample_data()
{
EnterCriticalSection(&hLock);
m_global = rand();
if ((m_global % 0xC5F) == 0)
{
// handle error
return;
}
LeaveCriticalSection(&hLock);
}

Listing 8.7 Sampling Data with Error Checking

After making these changes and rebuilding, the application becomes

unstable. In most instances, the application runs without any problems.
However, in certain circumstances, the application stops printing data.
How do you determine whats going on?
The key to isolating the problem is capturing a trace of the sequence
of events that occurred prior to the system hanging. This can be done
 Chapter 8: Multi-threaded Debugging Techniques 231

with a custom trace buffer manager or with tracepoints. This example

uses the trace buffer implemented in Listing 8.1.
Now armed with a logging mechanism, you are ready to run the
program until the error case is triggered. Once the system fails, you can
stop the debugger and examine the state of the system. To do this, run
the application until the point of failure. Then, using the debugger, stop
the program from executing. At this point, youll be able bring up the
Threads window to see the state information for each thread, such as the
one shown in Figure 8.1.

Figure 8.1 Examining Thread State Information Using Visual Studio 2005

When you examine the state of the application, you can see that the
consumer thread is blocked, waiting for the process_data() call to
return. To see what occurred prior to this failure, access the trace buffer.
With the application stopped, call the PrintTraceBuffer() method
directly from Visual Studios debugger. The output of this call in this
sample run is shown in Figure 8.2.
232 Multi-Core Programming

1 Thread ID |Timestamp Msg

2 ---------|--------- |------------------------------------
3 0x0000728|1137395188|Producer: sampled data value: 0x29
4 0x00005a8|1137395188|Consumer: processed data value: 0x29
5 0x0000728|1137395189|Producer: sampled data value: 0x78
6 0x00005a8|1137395189|Consumer: processed data value: 0x78
7 0x0000728|1137395190|Producer: sampled data value: 0x18BE
8 0x0000728|1137395190|Producer: sampled data value: 0x6784
9 0x0000728|1137395190|Producer: sampled data value: 0x4AE1
10 0x0000728|1137395191|Producer: sampled data value: 0x3D6C

Figure 8.2 Output from trace buffer after Error Condition Occurs

Examination of the trace buffer log shows that the producer thread is
still making forward progress. However, no data values after the first two
make it to the consumer. This coupled with the fact that the thread state
for the consumer thread indicates that the thread is stuck, points to an
error where the critical section is not properly released. Upon closer
inspection, it appears that the data value in line 7 of the trace buffer log
is an error value. This leads up back to your new handling code, which
handles the error but forgets to release the mutex. This causes the
consumer thread to be blocked indefinitely, which leads to the consumer
thread being starved. Technically this isnt a deadlock situation, as the
producer thread is not waiting on a resource that the consumer thread
holds.
The complete data acquisition sample application is provided on this
books Web site, www.intel.com/intelpress/mcp.

Multi-threaded Debugging Using GDB

For POSIX threads, debugging is generally accomplished using the GNU
Project Debugger (GDB). GDB provides a number of capabilities for
debugging threads, including:
Automatic notification when new threads are created
Listing of all threads in the system
Thread-specific breakpoints
The ability to switch between threads
The ability to apply commands to a group of threads
 Chapter 8: Multi-threaded Debugging Techniques 233

Not all GDB implementations support all of the features outlined here.
Please refer to your systems manual pages for a complete list of
supported features.

Notification on Thread Creation

When GDB detects that a new thread is created, it displays a message
specifying the threads identification on the current system. This
identification, known as the systag, varies from platform to platform.
Here is an example of this notification:
Starting program: /home/user/threads
[Thread debugging using libthread_db enabled]
[New Thread -151132480 (LWP 4445)]
[New Thread -151135312 (LWP 4446)]

Keep in mind that the systag is the operating systems identification for a
thread, not GDBs. GDB assigns each thread a unique number that
identifies it for debugging purposes.

Getting a List of All Threads in the Application

GDB provides the generic info command to get a wide variety of
information about the program being debugged. It is no surprise that a
subcommand of info would be info threads. This command prints a list
of threads running in the system:
(gdb) info threads
2 Thread -151135312 (LWP 4448) 0x00905f80 in vfprintf ()
from /lib/tls/libc.so.6
* 1 Thread -151132480 (LWP 4447) main () at threads.c:27
The info threads command displays a table that lists three properties
of the threads in the system: the thread number attached to the thread by
GDB, the systag value, and the current stack frame for the current thread.
The currently active thread is denoted by GDB with the * symbol. The
thread number is used in all other commands in GDB.

Setting Thread-specific Breakpoints

GDB allows users that are debugging multi-threaded applications to
choose whether or not to set a breakpoint on all threads or on a
particular thread. The much like the info command, this capability is
enabled via an extended parameter thats specified in the break
command. The general form of this instruction is:
break linespec thread threadnum
234 Multi-Core Programming

where linespec is the standard gdb syntax for specifying a breakpoint,

and threadnum is the thread number obtained from the info threads
command. If the thread threadnum arguments are omitted, the
breakpoint applies to all threads in your program. Thread-specific
breakpoints can be combined with conditional breakpoints:
(gdb) break buffer.c:33 thread 7 if level > watermark
Note that stopping on a breakpoint stops all threads in your program.
Generally speaking this is a desirable effectit allows a developer to
examine the entire state of an application, and the ability to switch the
current thread. These are good things.
Developers should keep certain behaviors in mind, however, when
using breakpoints from within GDB. The first issue is related to how
system calls behave when they are interrupted by the debugger. To
illustrate this point, consider a system with two threads. The first thread
is in the middle of a system call when the second thread reaches a
breakpoint. When the breakpoint is triggered, the system call may return
early. The reasonGDB uses signals to manage breakpoints. The signal
may cause a system call to return prematurely. To illustrate this point,
lets say that thread 1 was executing the system call sleep(30). When
the breakpoint in thread 2 is hit, the sleep call will return, regardless of
how long the thread has actually slept. To avoid unexpected behavior
due to system calls returning prematurely, it is advisable that you check
the return values of all system calls and handle this case. In this example,
sleep() returns the number of seconds left to sleep. This call can be
placed inside of a loop to guarantee that the sleep has occurred for the
amount of time specified. This is shown in Listing 8.8.

int sleep_duration = 30;

do
{
sleep_duration = sleep(sleep_duration);
} while (sleep_duration > 0);

Listing 8.8 Proper Error Handling of System Calls

The second point to keep in mind is that GDB does not single step all
threads in lockstep. Therefore, when single-stepping a line of code in one
thread, you may end up executing a lot of code in other threads prior to
returning to the thread that you are debugging. If you have breakpoints
 Chapter 8: Multi-threaded Debugging Techniques 235

in other threads, you may suddenly jump to those code sections. On

some OSs, GDB supports a scheduler locking mode via the set scheduler-
locking command. This allows a developer to specify that the current
thread is the only thread that should be allowed to run.

Switching between Threads

In GDB, the thread command may be used to switch between threads. It
takes a single parameter, the thread number returned by the info threads
command. Here is an example of the thread command:
(gdb) thread 2
[Switching to thread 2 (Thread -151135312 (LWP 4549))]#0
PrintThreads (num=0xf6fddbb0) at threads.c:39
39 {
(gdb) info threads
* 2 Thread -151135312 (LWP 4549) PrintThreads (num=0xf6fddbb0)
at threads.c:39
1 Thread -151132480 (LWP 4548) main () at threads.c:27
(gdb)

In this example, the thread command makes thread number 2 the active
thread.

Applying a Command to a Group of Threads

The thread command supports a single subcommand apply that can be
used to apply a command to one or more threads in the application. The
thread numbers can be supplied individually, or the special keyword all
may be used to apply the command to all threads in the process, as
illustrated in the following example:
(gdb) thread apply all bt
Thread 2 (Thread -151135312 (LWP 4549)):
#0 PrintThreads (num=0xf6fddbb0) at threads.c:39
#1 0x00b001d5 in start_thread () from
/lib/tls/libpthread.so.0
#2 0x009912da in clone () from /lib/tls/libc.so.6

Thread 1 (Thread -151132480 (LWP 4548)):

#0 main () at threads.c:27
39 {
(gdb)

The GDB backtrace (bt) command is applied to all threads in the system.
In this scenario, this command is functionally equivalent to: thread
apply 2 1 bt.
236 Multi-Core Programming

Key Points
This chapter described a number of general purpose debugging
techniques for multi-threaded applications. The important points to
remember from this chapter are:
Proper software engineering principles should be followed when
writing and developing robust multi-threaded applications.
When trying to isolate a bug in a multi-threaded application, it is
useful to have a log of the different sequence of events that led
up to failure. A trace buffer is a simple mechanism that allows
programmers to store this event information.
Bracket events that are logged in the trace buffer with before
and after messages to determine the order in which the events
occurred.
Running the application in the debugger may alter the timing
conditions of your runtime application, masking potential race
conditions in your application.
Tracepoints can be a useful way to log or record the sequence of
events as they occur.
For advanced debugging, consider using the Intel software tools,
specifically, the Intel Debugger, the Intel Thread Checker, and
the Intel Thread Profiler.
 Chapter 9
Single-Core
Processor
Fundamentals
o gain a better understanding of threading in multi-core hardware, it
T is best to review the fundamentals of how single-core processors
operate. During the debugging, tracing, and performance analysis of
some types of programs, knowing a processors details is a necessity
rather than an option. This chapter and Chapter 10 provide the
architectural concepts of processors that are pertinent to an
understanding of multi-threaded programming. For internal instruction-
level details, you should consult Intel Software Developers Guides at
Intels Web site.
This chapter discusses single-core processors as a basis for
understanding processor architecture. If you are already familiar with the
basics of processors and chipsets, you might skip this chapter and move
directly to Chapter 10.

Processor Architecture Fundamentals

The term processor has become loosely defined. A more precise
definition is developed in the following sections. A chipset is the set
of chips that helps processors interact with physical memory and other
components. Here, the chip is actually a processor but without
centralized main processing capability. A block diagram with all the basic
components in a computer system is represented in Figure 9.1.

237
238 Multi-Core Programming

Motherboard

Processor

Add-On Physical Memory

Cards
Chipset
(Must have
(Optional) minimum)

Connect with all

I/O cevices

Figure 9.1 Basic Components in a Computer System

You might be familiar with two chips in the chipset. Previously these
chips were known as the Northbridge and Southbridge and they were
connected by a shared PCI bus. Intel changed the implementation and
started using dedicated point-to-point connections or direct media
interface (DMI) between these two chips and introduced Intel Hub
Architecture (IHA), as shown in Figure 9.2. IHA replaced the Northbridge
and Southbridge with the Memory Controller Hub (MCH) and the I/O
Controller Hub (ICH). When graphics and video features are built into
the MCH, it is called the Graphics Memory Controller Hub (GMCH). A
front side bus (FSB) attaches the chipset to the main processor.
To understand the impact of the hardware platform on an
application, the questions to pose are which processor is being used,
how much memory is present, what is the FSB of the system, what is the
cache size, and how the I/O operations take place? The answer to most
of these questions is dictated by the processor.
The smallest unit of work in a processor is handled by a single
transistor. A combination of transistors forms a logic block and a set of
logic blocks create a functional unitsome examples are the Arithmetic
Logic Unit (ALU), Control Units, and Prefetch Units. These functional
units receive instructions to carry out operations. Some functional units
 Chapter 9: Single-Core Processor Fundamentals 239

are more influential than others and some remain auxiliary. The
functional units, or blocks, form a microprocessor or Central Processing
Unit (CPU). A high-level block diagram of a microprocessor is shown in
Figure 9.3(a). The manufacturing process of a microprocessor produces a
physical die and the packaged die is called the processor. Figure 9.3(b)
shows a photo of a die. Inside a computer system, the processor sits on a
socket. To show the physical entity of processor and socket, see
Figure 9.3(c). Sometimes the processor is referred to as the CPU. For
simplicitys sake, this book refers to processor and microprocessor
interchangeably. Different processors usually have a different number of
functional units.

Motherboard

Processor

Chipset
Add-On MCH Physical Memory
Graphics (Memory (Must have
Controller
(Optional) Hub) minimum)
DMI
(Direct Media
Interface)
Add-On ICH
Cards (IO Controler
(Optional) Hub)

Connect with all

I/O cevices

Figure 9.2 A System Showing MCH, ICH, and FSB

240 Multi-Core Programming

Processor

On-Die Cache
L0 . . . Ln

Local APIC
Control Logic
(Advanced Programmable
Interrupt Controller)

Interace Unit
Execution
Resources
(ALU, Control Unit)
Register Array

(a) Functional Blocks Representation (b) Die after Production from a

of a Microprocessor, where Lmax = 3 Processed Silicon Wafer

Heat-Sink

Processor
(Single-Core or
Multi-Core)

Single
Socket

(c) Processor with Other Components

Figure 9.3 Processor, Die, and Socket

A processor fetches software instructions as input, performs

instruction decode operations to make instructions understood by the
processor, does some specific tasks, and finally produces the output, as
illustrated in Figure 9.4. All these operations are done through the
functional blocks inside a processor and all of the pipeline stages are
within the boundary of a processor.
 Chapter 9: Single-Core Processor Fundamentals 241

System Bus or FSB (Front Side Bus)

After operations
Fetch Functional
inside a processor
Block gets
results get back to
instructions from
main memory
main memory during
during post
fetch cycle
execution cycle

On-Die Cache
L0 . . . Ln

Local APIC
Control Logic Decode and
(Advanced Programmable
Interrupt Controller) execute
operations
Interace Unit
Execution
Resources
(ALU, Control Unit)
Register Array

Processor

Figure 9.4 Processor Attached with the System Bus Showing Basic
Operational Steps

Now lets review the internals of a processor.

The on-die caches are usually referred to as levels: L1, L2, and
L3. L1 is the smallest and L3 is the largest. Most of the 32-bit
processors do not yet have an L3 cache, whereas the currently

available Intel Itanium processors have large L3 caches, such
as the 6-megabyte L3 cache in the Itanium 2 processor.
The Local Advanced Programmable Interrupt Controller
(Local APIC) unit is specific to a processor and provides
interrupt handling capability to a specific processor. This is
242 Multi-Core Programming

not the I/O APIC. The I/O APIC is a part of the chipset that
supports interrupt handling of different I/O devices through
the Local APIC. The I/O APIC is an off-chip unit and usually a
part of a multi-processor-based chipset.
The interface unit is the functional block that helps to
interface a processor with the system bus or front side bus
(FSB).
The register array is the set of registers present in a processor.
The number of registers can vary significantly from one
generation of processor to another: 32-bit processors without
Intel Extended Memory 64 Technology (Intel EM64T) have

only eight integer registers, whereas 64-bit Itanium
processors have 128 integer registers.
The execution resources include the integer ALU, Floating-
Point execution, and branch units. The number of execution
units and the number of active execution units per cycle
referred to as the number of issue portsvary from processor
to processor. The execution speed of functional blocks varies
as well and these implementations get improved from
generation to generation. The number of execution units is
sometimes referred to as the machine width. For example, if
the processor has six execution units, the processor is said to
be a six-wide machine.
Other types of functional blocks are available in the processor and they
vary with respect to the type of processor as well. There are areas in a
processor referred to as queues that temporarily retain instructions prior
to going into the next phase of operation through the pipeline stages.
The scheduler is another functional block. It determines when micro-
operations are ready to execute based on the readiness of their
dependent input register operand sources and the availability of the
execution resources the micro-operations need to complete their
operation.
The execution flow of operations in a processor is shown in
Figures 9.5 and 9.6. These figures depict the basic four steps of the
pipeline: fetch, decode, execute, and write. In reality the process is
somewhat more complicated.
 Chapter 9: Single-Core Processor Fundamentals 243

System Bus or FSB (Front Side Bus)

Bus Interface Unit

Fetch / Decode

Perform some specific tasks

and produce results

Processor

Figure 9.5 Basic Execution Flow in a Processor

Static Program

Instruction
reorder
and
Window of Execution commit
Instruction Dispatch

Instruction
Instruction Issue

fetch and
branch
prediction

Figure 9.6 Basic Execution Flow in a Superscalar Processor

244 Multi-Core Programming

Different types of Instruction Set Architecture (ISA) processors exist,

but in reality the basic aspects of a processor core remain the same. Intel
1
technology can be divided into two primary types of cores: superscalar
and Explicitly Parallel Instruction Computing (EPIC). All the processors
discussed here are based on these two types. You might already be
familiar with a superscalar core. Intels mainstream superscalar processor

architecture began in 1993 with the introduction of the Intel Pentium

processor. This superscalar architecture evolved into the Intel NetBurst
microarchitecture, as found in the Pentium 4 processor.
Figure 9.7 provides a block diagram of the Intel Pentium 4 processor.
Here the functional blocks are partitioned into three distinctive segments,
front end, back end or execution core, and memory subsystem.

System Bus or FSB (Front Side Bus)

Processor

Bus Interface Unit

Memory Subsystem

L2 Cache L1 Cache

Integer and Floating

Front End
Point Execution Units

Fetch / Trace Cache

Decode Microcode ROM Out-of-Order Retirement
(OOO)
Execution
Logic
BTB / Branch Branch History Update
Prediction
Back End or Execution Core

Figure 9.7 Execution Flow inside Intel Pentium 4 Processor

1
A processor with a single pipeline is called a scalar processor and a CPU with multiple pipelines is
called a superscalar processor.
 Chapter 9: Single-Core Processor Fundamentals 245

Comparing Superscalar and EPIC Architecture

Intel recently started developing multi-core processors for both
superscalar and EPIC architectures. The superscalar architecture is
commonly referred to as wide, speculative, and dynamic in nature.
To provide a better understanding of differences between these two
architectures, Table 9.1 compares superscalar and EPIC architecture
and Figure 9.8 shows the operational flow in these two
architectures.

Table 9.1 Comparison of Superscalar and EPIC Architecture

Superscalar EPIC
Supports 32-bit and 64-bit (Intel EM64T) Supports 64-bit (Intel Itanium
Architecture)
Effective resource utilization with Massive resources with large number of
minimum number of registers array registers array
RISC-like instructions RISC-like instructions bundled into
groups of three
Has multiple parallel execution units Has multiple parallel execution
units
Runtime scheduling Mostly static scheduling with the help of
compiler
Single path speculative execution with Both paths of speculative execution with
branch prediction branch prediction
Loads data from memory only when Speculatively loads data before its
needed, and tries to find the data in the needed, and still tries to find data in the
caches first caches first
246 Multi-Core Programming

Superscaler Architecture EPIC

Original Parallel
Original Source
Source Hardware Machine Code
Code Itanium 2
Code based
Implicity
Compiler Parallel compiler

Sequential
Machine Code

Multiple execution
efficiently but units used
Execution Units difficult more efficiently but Massive
to use efficiently compiler more complex Resources

Figure 9.8 Operational Comparison between Superscalar and EPIC

Key Points
Understanding the basics of a single-core processor is essential to
comprehend how threading works on a multi-core processor. The
important concepts and terms to keep in mind are:
There are different functional blocks that form a microprocessor
such as, Arithmetic Logic Unit, Control Units, and Prefetch Units.
A chipset is used to interface the processor to physical memory
and other components.
A processor is the container of the dies, and the die is the
microprocessor or CPU. In loose terms, processor and
microprocessor get used interchangeably.
The high-level operations for multi-core processors remain the
same as for single-core processors.
Two fundamentally different generic architectures are available
from Intel: wide superscalar and EPIC.
Now that the basic building blocks of a processor have been covered, the
following chapter explores multi-core processor architecture from a
hardware perspective, focusing on the Pentium 4 processor and Itanium
architecture.
 Chapter 10

Threading on Intel
Multi-Core
Processors
he concepts of threading from a software perspective were covered
T in previous chapters. Chapter 2 also touched briefly on threading
inside hardware and Chapter 9 covered the concepts of single-core
processors. This chapter describes in more detail what threading inside
hardware really means, specifically inside the processor. Understanding
hardware threading is important for those developers whose software
implementation closely interacts with hardware and who have control
over the execution flow of the underlying instructions. The degree to
which a developer must understand hardware details varies. This chapter
covers the details of the multi-core architecture on Intel processors for
software developers, providing the details of hardware internals from a
threading point of view.

Hardware-based Threading
Chapter 9 describes the basics of the single-core processor. In most
cases, threaded applications use this single-core multiple-issue
superscalar processor. The threading illusion materializes from the
processor and that is called instruction level parallelism (ILP). This is
done through a context-switch operation. The operational overhead of
context switching should be limited to a few processor cycles. To
perform a context switch operation, the processor must preserve the
current processor state of the current instruction before switching to

247
248 Multi-Core Programming

other instruction. A processor keeps ongoing operational information

mainly in registers and a policy dictates this context switch operation.
The simplest class of processor is single-issue, single-thread (SIST) or
single-threaded scalar-based processor. For SIST, the OS handles multiple
threads. In terms of hardware resource sharing and the level of
granularity of resource hold time, there are two types of processors
available: coarse-grained multi-threading (CGMT) and fine-grained multi-
threading (FGMT). Each maintains a policy of sharing resources among
threads.
For coarse-grained multi-threading, or switch-on-event multi-
threading, a thread has full control over processor resources for a
specified quantum of time or number of processor cycles. In fine-grained
multi-threading, the thread switching takes place at an instruction-cycle
boundary. That means the overhead associated with thread switching is
higher for a coarse-grained than for a fine-grained implementation. To
reflect the policy associated with these processors, coarse-grained multi-
threading is also referred to as blocked multi-threading scalar, and fine-
grained multi-threading as interleaved multi-threading scalar, illustrated in
Figure 10.1. Both fine- and coarse-grained multi-threading are sometimes
referred to as temporal multi-threading (TMT). Chunky Context Switch
Granular Context Switch
Cycle or Pipeline Stages

.. .. ..
. . .

Unused Cycle
Fine-grained Coarse-grained
Single-issue,
Multi-threading Multi-threading
Single-thread
or Interleaved or Blocked
or Single-threaded
Multi-threading Multi-threading
Superscaler
Superscalar Superscalar

Figure 10.1 Different Threading Scenarios on Multi-issue Processors

 Chapter 10: Threading on Intel Multi-Core Processors 249

From an application level, it appears that multiple threads run at the

same time, but in reality the system does not have enough resources to
support those threads simultaneously. The OS scheduler, in combination
with the hardware scheduler and execution core, gives the impression of
threading.
For systems with multiple processors or symmetric multi-processor
(SMP) systems with shared memory, the scenario is different. In an SMP
environment, the system can utilize thread-level parallelism (TLP) as well
by running different threads in parallel on different processors. The OS
scheduler is also responsible for handling this thread balancing act on the
system. The use of ILP and TLP lack the benefit of resource utilization. To
address the issue of processor resource utilization, the introduction of
simultaneous multi-threading (SMT) allows multiple threads to compete
for shared available resources, shown in Figure 10.2.
No thread switching, both threads
compete for available resources
Cycle or Pipeline Stages

.. ..
. . Unused Cycle

Figure 10.2 SMT Handles Multiple Threads

SMT hardware is effective for those applications that require

complementary hardware resources during runtime. A multi-threaded
singleton applicationan application that has dependency on a specific
functional unit of the CPU, such as integer or floating point functional
unitsmight suffer a performance penalty on an SMT platform. In an
SMT processor, TLP gets converted into ILP and accommodates variations
among ILP and TLP. In terms of granularity, to utilize resources
effectively an SMT processor exploits both coarse-grained and fine-
granted parallelism.
250 Multi-Core Programming

When a processor has two or more cores, then that processor is

referred to as chip multiprocessing (CMP). Here, each core executes
hardware threads independently of other hardware threads, and shared
memory helps to maintain inter-thread communication, shown in
Figure 10.3. This independent thread execution on a multi-core
processor is referred as chip multi-threading (CMT).

No thread switching, use distinctive

Cycle or Pipeline Stages

individual execution path

.. ..
. . Unused Cycle

Figure 10.3 CMP Handles Multiple Threads

The concept of CMP has been around for a while in the specialized
processor domain, in areas like communication and DSP. In CMP
technology, multiple processors reside on a single die. To extend the
usability of the CMP in the general processor domain, the industry
introduced the concept of multi-core processors, which are slightly
different than CMPs even though many publications started using CMP as
a generic description of multi-core processors. In CMPs, multiple
processors get packed on a single die, whereas for multi-core processors,
a single die contains multiple cores rather than multiple processors. The
concept of CMP can be confused with the existence of multiprocessor
systems-on-chip (MPSoC). CMP and MPSoC are two different types of
processors used for two different purposes. CMP is for general-purpose
ISA hardware solutions, whereas MPSoC is used in custom architectures.
In simple terms, the CMP is the one that has two or more conventional
processors on a single die, as shown in Figure 10.4.
 Chapter 10: Threading on Intel Multi-Core Processors 251

1 MB L 2I 2 Way
Multi-threading

Processor
Core

L3 Tag
L2 Cache Dual
(256K) core
L3 Cache

Bus
Logic

2 x 12 MB L 3 Arbiter
I/O caches

(a) Die of a Single-core Processor (b) Die of a Dual-Core Processor

Figure 10.4 Single-core and Dual-Core Processor Dies

Several dimensions of technological evolution influenced the

development of hardware threading. Process technology helps to
manufacture smaller transistors and accommodates more transistors in a
smaller package and keeps everything within the thermal envelopethe
amount of heat allowed for a single processor. When you think of a
processor, you must realize that there can be a good number of

instructions in flight at operational time. In the Intel NetBurst
microarchitecture as many as 126 instructions remain in flight at any one
time, positioned in various stages of execution and ready to execute
simultaneously. To handle these many instructions and utilize processor
resources, it is essential to incorporate parallelism effectively in a
processor. This is one of the major reasons for the evolution of
processors from superscalar to SMT to multi-core architecture.

Threading from Intel

Now that you have an idea what hardware threading means, you can
easily guess that Intel has been implementing threading in processors for

some timein fact with the introduction of the Intel Pentium
superscalar processor in 1993. The performance was not impressive
compared to the current standard, but was just a beginning of threading-
based solutions on a processor. The progress continued and the next
shift took place in 2000 with the introduction of Hyper-Threading
Technology (HT Technology) for the 32-bit world and by the addition of
Explicit Parallel Instruction Computing (EPIC) architecture with the
252 Multi-Core Programming

launch of 64-bit Itanium processors. The next wave from Intel came
with the addition of dual-core processors in 2005, and further
developments are in the works. To understand Intel threading solutions
from the processor level and the availability of systems based on these

processors, review the features of brands like Intel Core Duo, Intel

Pentium Processor Extreme Edition, Intel Pentium D, Intel Xeon , Intel
Pentium 4, and Intel Itanium 2. As stated before, when you are going to
select different types of processor for your solution, you have to make
sure the processor is compatible with the chipset. To learn more details
on processors and compatibility, visit the Intel Web site.

Hyper-Threading Technology
Hyper-Threading Technology (HT Technology) is a hardware mechanism
where multiple independent hardware threads get to execute in a single
cycle on a single superscalar processor core, as shown in Figure 10.5.
The implementation of HT Technology was the first SMT solution for
general processors from Intel. In the current generation of Pentium 4
processors, only two threads run on a single-core by sharing, and
replicating processor resources.

Physical Package Physical Package

Architectural Logical Logical
State Processor 0 Processor 1
Execution Architectural Architectural
Engine State State

Local APIC Execution Engine

Local APIC Local APIC
Bus Interface
Bus Interface

System Bus System Bus

Single Processor Single Processor with
Hyper-Threading Technology

Figure 10.5 Single Processor System without Hyper-Threading Technology and

Single Processor System with Hyper-Threading Technology
 Chapter 10: Threading on Intel Multi-Core Processors 253

Inside a processor with HT Technology, two threads share resources

from a single core, and that is why these threads are referred to as logical
processors. In terms of physical processor core resources, a Pentium 4
processor with HT Technology and one without are almost the same.
Only the die size is increased for the additional logic on the processor
with HT Technology. The number of registers on processors with and
without HT Technology remains the same. Obviously only one of these
two threads can use a shared resource at a time.
From the OS perspective, the system represents two logical
processors. This configuration allows a thread to be executed on each
logical processor. Instructions from both threads are simultaneously
dispatched for execution by the processor core. The processor core
executes these two threads concurrently, using out-of-order instruction
scheduling to keep execution units as busy as possible during each clock
cycle. Figure 10.6 shows that the time taken to process n threads on a
single processor is significantly more than a single-processor system with
HT Technology enabled. This is because with HT Technology enabled,
two logical processors process two threads concurrently on one physical
processor.

Multithreading Hyper-Threading Technology

App 0 App 1 App 2 App 0 App 1 App 2

T0 T1 T2 T3 T4 T5 T0 T1 T2 T3 T4 T5
Thread Pool Thread Pool

CPU
CPU

LP0
CPU T0 T1 T2 T3 T4 T5 T0 T2 T4
LP1 T1 T3 T5
Time CPU
2 Threads per
Processor Time

Figure 10.6 Multi-threaded Processing using Hyper-Threading Technology

254 Multi-Core Programming

Difference between Multiprocessor and Hyper-Threading Technology

Multiprocessor technology is referred to as MP. In MP multiple physical
processors exist, whereas HT Technology relates to only one physical
processor. In an MP environment, each processor could be enabled with
HT Technology as well, as shown in Figure 10.7. For an MP environment
without HT Technology, each thread dynamically gets a fixed number of
devoted functional blocks in a processor, whereas with HT Technology
the resources get shared among threads and a thread assignment policy
determines the resource utilization.

MP HT MP
Physical Package Physical Package Physical Package Physical Package
Logical Logical Logical Logical
Processor 0 Processor 1
Architectural Architectural
Processor 0 Processor 1 State State
Architectural Architectural Architectural Architectural
State State State State Execution Execution
Execution Engine Execution Engine Engine Engine
Local APIC Local APIC Local APIC Local APIC Local APIC Local APIC
Bus Interface Bus Interface Bus Interface Bus Interface

System Bus System Bus

Figure 10.7 Multiprocessor with and without Hyper-Threading Technology

Hyper-Threading Technology Architecture

HT Technology is integrated into the Intel NetBurst microarchitecture
using 90nm technology. The operational protocols and algorithms are
improved to accommodate two execution flows of hardware threads.
The various generations of processors with HT Technology are enhanced
by additional features, whereas the architectural core remains the same.
Figure 10.8 shows the different functional blocks in the architecture of
the Pentium 4 processor with HT Technology.
The instruction decoding phase is independent of the execution
phase. This separation helps to maintain the flow of the two threads. The
instruction fetch logic keeps two streaming buffers for use with both
threads and two instruction pointers (IP) to track the progress of
instruction fetches for the two logical processors. In the case of branch
prediction, branch prediction structures return a stack buffer and branch
history buffer that get duplicated, and a large global history array is
 Chapter 10: Threading on Intel Multi-Core Processors 255

shared with entries that are tagged with logical processor IDs. The
decode logic preserves two copies of all the necessary states required to
perform an instruction decode, even though the decoding operations are
done through a coarse-grained scheme in a single cycle.

ITLB/Prefetcher Front-End BTB

Decoder

Microcode
Rom
BTB
Execution Trace Cache
Micro-op
Queue
L2 Cache and Control
FSB Bus Interface

Allocator/Register Renamer
System Bus

Memory micro-op Queue Integer/FP micro-op Queue

Memory Fast FP Slow/General FP Simple FP

Scheduler Scheduler Scheduler Scheduler

Integer Register File FP Register

Load Store 2x ALU 2x ALU Slow ALU FP, MMX FP

Address Address Simple Simple Complex SSE, SSE2 Move
AGU AGU Inst. Inst. Inst. SSE3 FXCH

L1 Data Cache and DTLB

ITLB: Instruction Translation Lookaside Buffer AGU: Address Generation Unit BTB:Branch Target Buffer
DTLB: Data Translation Lookaside Buffer FPx: Floating Point FP, MMX, SSE, SSE2, SSE3 FPm: Floating Point Move, FXCH

Figure 10.8 Pentium 4 Processor Architecture with Queue and Scheduler

Splits

The decode logic passes decoded instructions to the trace cache, also
referred to as the advanced instruction cache. In reality, this is somewhat
different than the conventional instruction cache. The trace cache stores
already-decoded instructions in the form of micro-ops and maintains data
integrity by associating a logical processor ID. The inclusion of the trace
cache helps to remove the complicated decodes logic from the main
execution phase. The trace cache orders the decoded micro-ops into
program-ordered sequences or traces. If both hardware threads need
256 Multi-Core Programming

access to the trace cache, the trace cache provides access with a fine-
grained approach rather than coarse-grained. The trace cache can hold
up to 12K micro-ops, but not every required instruction can reside in the
trace cache. That is why, when the requested instruction micro-op is not
available in the trace cache, the instruction needs to bring it from L2
cachethis event is called a Trace Cache Miss. On the other hand, when
the instruction micro-ops remain available in trace cache and instruction
flow does not need to take extra steps to get required instructions from
L2, the event is referred to as a Trace Cache Hit.
In the front end of a processor with HT Technology, both
hardware threads make independent progress and keep data
association. The micro-op queues decouple the front end from the
execution core and have a hard partition to accommodate two
hardware threads. Once the front end is ready to prepare the
microcode, the operational phase gets transferred to the backend out-
of-order execution core, where appropriate execution parallelism
takes place among microcode streams. This is done with the help of
distributor micro-op queues and schedulers which keep the correct
execution semantics of the program. To maintain the two hardware
threads register resource allocation, two Register Allocation Tables
(RATs) support two threads. The register renaming operation is done
in parallel to allocator logic. The execution is done by the advanced
dynamic execution engine (DEE) and the rapid execution engine
(REE). Six micro-ops get dispatched in each cycle through DEE and
certain instructions are executed in each half cycle by REE. When two
hardware threads want to utilize back-end, each thread gets allocation
through a fine-grained scheme and a policy is established to limit the
number of active entries each hardware thread can have in each
scheduler queue. To provide ready micro-ops for different ports, the
collective dispatch bandwidth across all of the schedulers is twice the
number of micro-ops received by the out-of-order core.
Once the out-of-order execution core allows instructions from both
threads interleaved in an arbitrary fashion to complete execution, it
places issued micro-ops in the reorder buffer by alternating between two
hardware threads. If for some instruction, one hardware thread is not
ready to retire micro-ops, other threads can utilize the full retirement
bandwidth.
 Chapter 10: Threading on Intel Multi-Core Processors 257

In the memory subsystem, the Data Translation Lookaside Buffer

(DTLB) is a shared resource but maintains hardware thread tags or logical
processor tags to maintain data integrity. The rest of the cache
hierarchies get shared by hardware threads. Inside the bus, no priority is
assigned to logical processors or hardware threads, even though the
distinction between requests from two logical processors is maintained
reliably. The interrupt maintenance is done through local APICs, which
are unique to each logical processor.

Multi-Core Processors
To understand multi-core processors, this section extends the concepts
of single core and differentiates the meaning of core from that of
processor. The following sections also cover the basics of the multi-core
architecture, what is available today, and what may be available beyond
multi-core architecture.

Architectural Details
Chapter 9 reviewed how a single-core processor contains functional
blocks, where most of the functional blocks perform some specific
operations to execute instructions. The core in this case is a combination
of all of the functional blocks that directly participate in executing
instructions. The unified on-die Last Level Cache (LLC) and Front Side
Bus (FSB) interface unit could be either part of the core or not,
depending on the configuration of the processor.
Some documents exclusively differentiate between core and
execution core. The only difference is that an execution core is the main
set of functional blocks that directly participate in an execution, whereas
core encompasses all available functional blocks. To remain consistent,
this book tries to distinguish the differences. In Figure 10.9, different
core configurations are shown. Using shared LLC in a multi-core
processor, the cache coherency complexity is reduced, but there needs
to be a mechanism by which the cache line keeps some identifying tag
for core association or dynamically splits the cache for all cores. Also,
when the FSB interface gets shared, this helps to minimize FSB traffic.
Proper utilization of a multi-core processor also comes from a compatible
chipset.
258 Multi-Core Programming

Processor

Execution Core
(EC)

Last Level Cache

(LLC)

FSB Bus
Interface Unit

System Bus or FSB (Front Side Bus)

(a) Single Core Processor

Processor Processor

Execution Core Execution Core Execution Core Execution Core

(EC) (EC) (EC) (EC)

Last Level Cache Last Level Cache Last Level Cache Last Level Cache
(LLC) (LLC) (LLC) (LLC)
FSB Bus Interface Unit FSB Bus Interface Unit FSB Bus Interface Unit FSB Bus Interface Unit

System Bus or FSB (Front Side Bus) System Bus or FSB (Front Side Bus)

(b) Multi-Core Processor with Two (c) Actual representation of (b)

Cores and Individual FSB to show shared FSB
Processor Processor

Execution Core Execution Core Execution Core Execution Core

(EC) (EC) (EC) (EC)

Last Level Cache Last Level Cache Last Level Cache

(LLC) (LLC) (LLC)

FSB Bus Interface Unit FSB Bus Interface Unit

System Bus or FSB (Front Side Bus) System Bus or FSB (Front Side Bus)

(d) Multi-Core Processor with Two (e) Multi-Core Processor with Two
Cores and Shared FSB Cores and Shared LLC and FSB

Figure 10.9 Processor Core Configurations

 Chapter 10: Threading on Intel Multi-Core Processors 259

The number of cores can vary, but the cores remain symmetrical; that
is why you see product announcements for two, four, eight, or more
cores in processors. You will be seeing the representation of the number
n
of cores by 2 (where, in theory, 0 < n < ). Projected theoretical
representations always remain blocked by available technologies. With
the constraint in current technology, the proposed geometry of current
and upcoming multi-core processors is shown in Table 10.2.

Table 10.2 Disclosed Multi-Core Processors with Specific Features

Processor Brand Number LLC Size HT FSB Interface Unit
or Code Name of Technology (Shared or
Cores Present Independent)
Intel Core Duo 2 1 2 MB No Shared
2 1 MB

Intel Pentium D 2 No Independent
2 1 MB

Intel Pentium 2 Yes Independent
Processor Extreme
Edition
Intel codename 2 2 2 MB No Independent
Presler
Intel codename 2 2 x 2 MB Yes Independent
Dempsey
Intel codename 2 2 x 2 MB Yes Shared
Paxville
Intel Itanium 2 2 12 MB Yes Shared
processor
codenamed
Montecito

Table 10.3 shows only two physical cores. The number of threads
supported by these processors is currently limited to two cores, but with
respect to the platform, the number of threads varies with respect to the
chipset where these processors are being used. If the chipset supports N
number of processors, then the number of hardware threads for that
platform can be as high as N 2 2. For the latest updates about
available processors, visit the Intel Web site.
260 Multi-Core Programming

Networking Model of Multi-Core Processors

How do multiple cores communicate with each other and how are
these cores positioned inside a die? These concerns are similar to
network topology issues. The interconnection could be bus, mesh,
ring, cross-bar, and so on. Different vendors utilize different
topologies for these interconnections. Currently from Intel, the
interconnection follows the existing FSB-based connection scheme.
This approach has some important legacy aspects associated with it.
Remember, auxiliary and required components must support multi-
core processors and for that, all these components must support and
have cohesive features to handle multi-core processors. The core of
a multi-core processor does the same things that a single core based
processor does except that with a multi-core processor, cores have
to operate in a concerted way.
How about threads on these multi-core processors? What will
happen with your application as the number of cores increase? The
more cores that processors support, the more hardware threads you
get to utilize. The multi-core processor hardware is evolving with
updated protocols and improved logic. To implement threading in
software, you need to use a methodology for synchronization. If you
have ever performed operations to handle hardware threads directly,
you know the level of synchronization that needs to be done. The
layer above the hardware needs to support proper synchronization
of the processors. Otherwise, an application might not get the
expected performance gain. You also need to understand which
operating systems support these multi-core processors and which
compilers generate better code to deal with these many hardware
threads.

Comparison between Multiprocessors and Multi-Core Processors

A multiprocessor represents multiple physical processors, whereas a
multi-core processor represents multiple cores in a single physical
processor, as shown in Figure 10.10. That means a multiprocessor can be
built from a number of multi-core processors. Think of a multiprocessor
environment as containing multiple sockets where you can plug in
multiple processors. A multi-core processor resides on a single socket.
 Chapter 10: Threading on Intel Multi-Core Processors 261

(a) Multiprocessor with Single-Core Processors

(b) Multiprocessor with Multi-Core Processors

Figure 10.10 Multiprocessor with Single-Core and Multi-Core processors

Multi-Core for Itanium Architecture

The evolution of multi-core processors is not bounded by superscalar
architecture. Intel announced a next-generation multi-core processor into
the Explicitly Parallel Instruction Computing (EPIC) architecture domain
code-named Montecito. Montecito introduces two cores in a single die, as
shown in Figure 10.11 Even though the Itanium processor is an explicitly
parallel instruction computing processor which provides exclusive
instruction-level parallelism, to utilize resources and reduce miss
penalties, Montecito incorporates both ILP and TLP. The implementation
of ILP is similar to the implementation concepts of HT Technology,
where resources get shared by fine-grained and coarse-grained
parallelism, as well as SMT. In the current Itanium architecture, the
262 Multi-Core Programming

concept of HT Technology blends Switch-on Event Multi-threading

(SoEMT) for the cores and SMT for the memory hierarchy. The SoEMT is
a form of coarse-grained parallelism where the time constants are varied
based on dynamic behaviors of the instruction stream or events. Dont
confuse this term with Intel EM64T.

L1I Branch Instruction L1I Branch Instruction

Cache (16KB) Prediction TLB Cache (16KB) Prediction TLB

B B B I I M M M M F F B B B I I M M M M F F

Register Stack Engine / Re-name Register Stack Engine / Re-name

Branch & Floating Branch & Integer Floating

Predicate Integer Point Predicate Point
Registers Registers
Registers Registers Registers Registers

Branch Unit Integer Memory/ Floating Branch Unit Integer Memory/ Floating
Unit Integer Point Unit Unit Integer Point Unit

L1D Data L1D ALAT Data

ALAT Cache (16KB) TLB
Cache (16KB) TLB

L2D L2I L2D L2I

Cache (256KB) Cache (1MB) Cache (256KB) Cache (1MB)

Queues/ L3 Queues/ L3
Control Cache (12MB) Control Cache (12MB)
Arbiter

Synchronizer Synchronizer

System Interface

Figure 10.11 Multi-Core Architecture of Intel Itanium Processor code-named Montecito

The FSB interface is shared among the cores. Montecito supports two
cores in each socket and two hardware threads on each core. So, one
socket has four contexts. This can be seen as comparable to a dual-core
platform with HT Technology.
In Montecito, each core attaches to the FSB interface through the
arbiter, which provides a low-latency path for a core to initiate and
respond to system events while ensuring fairness and forward progress.
 Chapter 10: Threading on Intel Multi-Core Processors 263

The arbiter maintains communication with the core through a

synchronization functional block, as shown in Figure 10.12. The arbiter
maintains each cores unique identity to the FSB interface and operates at
a fixed ratio to the FSB interface frequency.

Processor
Processor
Execution Execution
Core Core
(EC) (EC)
Execution Core
Last Level Last Level
(EC)
Cache (LLC) Cache (LLC)

Last Level Cache L3 L3

(LLC) Synchronizer Synchronizer
L3

FSB Bus Arbitrater

Interface Unit
FSB Bus Interface Unit

System Bus or FSB (Front Side Bus) System Bus or FSB (Front Side Bus)

(a) Single Core Itanium 2 (b) Montecito Core
Processor

Figure 10.12 Single Core Itanium 2 Processor and Montecito Core

An asynchronous interface between the arbiter and each core enables

the core and cache frequency to vary as needed. This arbiter and the
synchronizers add a small amount of latency to transactions both from a
core to the system interface and from the system interface to a core.
The arbiter consists of a set of address queues, data queues,
synchronizers, control logic for core and FSB interface arbitration, error-
correction code (ECC) encoders/decoders, and parity generators. The
arbiter interleaves core requests on a one-to-one basis when both cores
have transactions to issue. When only one core has requests, it can issue
its requests without waiting for the other core to issue a transaction.
Because read latency is the greatest concern, the read requests are
typically the highest priority, followed by writes, and finally clean victim
notifications from the LLC.
264 Multi-Core Programming

Each core tracks the occupancy of the arbiters queues using a credit
system for flow control. As requests complete, the arbiter informs the
appropriate core of the type and number of de-allocated queue entries.
The cores use this information to determine which, if any, transaction to
issue to the arbiter. The arbiter manages the system interface protocols
while the cores track individual requests. The arbiter tracks all in-order
requests and maintains the system interface protocol. Deferred or out-of-
order transactions are tracked by the core with the arbiter simply passing
the appropriate system interface events on to the appropriate core. The
arbiter has the ability to support various legacy configurations by
adjusting where the agent identifiersocket, core, and/or threadis
driven on the system interface. The assignment of socket and core must
be made at power on and cannot be changed dynamically. The
assignment of a thread is fixed, but the driving of the thread identifier is
under Processor Abstraction Layer (PAL) control since it is for
information purposes only and is not needed for correctness or forward
progress.
In the core, one thread has exclusive access to the execution
resources (foreground thread) for a period of time while the other thread
is suspended (background thread). Thread control logic evaluates the
threads ability to make progress and may dynamically decrease the
foreground threads time quantum if it appears that it will make less
effective use of the core than the background thread. This ensures better
overall utilization of the core resources over strict temporal multi-
threading (TMT) and effectively hides the cost of long latency operations
such as memory accesses, especially the on-die LLC cache misses, which
has latency of 14 cycles. Other events, such as the time-out and forward
progress event, provide fairness, and switch hint events provide paths for
the software to influence thread switches. These events have an impact
on a threads urgency that indicates a threads ability to effectively use
core resources. Many switch events change a threads urgency, or the
prediction that a thread is likely to make good use of the core resources.
Each thread has an urgency value that is used as an indication of a
threads ability to make effective use of the core execution resources.
The urgency of the foreground thread is compared against the
background thread at every LLC event. If the urgency of the foreground
thread is lower than the background thread then the LLC event may
initiate a thread switch. Thread switches may be delayed from when the
control logic requests a switch to when the actual switch occurs. The
 Chapter 10: Threading on Intel Multi-Core Processors 265

reasons for delay include serialization operations and long latency

accesses. Urgency can take on values from 0 to 7. An urgency of 0
denotes that a thread has no useful work to perform. An urgency of 7 is
only used for a thread that is switched due to a time-out event when its
current urgency is 5. An external interrupt directed at the background
thread sets the urgency for the background thread at 6 to provide a
reasonable response time for interrupt servicing, but the urgency for the
current thread that receives an interrupt is not changed. The nominal
urgency is 5 and indicates that a thread is effectively using (or would
effectively use) the core execution resources (no LLC misses
outstanding). The urgency is reset to 5 when the background thread with
urgency above 5 becomes the foreground thread. Every LLC miss event
decrements the urgency by 1 after the urgency is compared, eventually
saturating at 0. Similarly, every LLC return event increments the urgency
by 1 before the urgency is compared saturating at 5.
Though most of the hardware threads are controlled by a processor
control functional block, in Montecito you would be able to control
threads using the hint@pause instruction. The hint@pause instruction is
used by software to initiate a thread switch. The intent is to allow code
to indicate that it does not have any useful work to do and that its
execution resources should be given to the other thread. Some later
event may change the work for the thread and should awaken the thread
such as an interrupt.
The hint@pause instruction forces a switch from the foreground
thread to the background thread. This instruction can be predicated to
conditionally initiate a thread switch. The current issue group retires
before the switch is initiated. Consequently, the following code
sequences are equivalent:
Hint at beginning of issue group:
hint@pause
add r1 = r2, r3
add r4 = r2, r0
Hint at end of issue group:
add r1 = r2, r3
add r4 = r2, r0
hint@pause
Having all these changes in Montecito does not affect legacy
software.
266 Multi-Core Programming

Multiple Processor Interaction

Up to this point, the book has covered hardware, architecture, and the
impact of software on a multi-core environment. As yet no detailed
information has been offered about the concepts of communication
among processors or cores. This section discusses details about multiple
processor intercommunication. It is impossible to organize a parallel
computational system without establishing communication channels or
other means of interaction between different parts of such a system.
Thus, completely isolating processors and preventing them from
exchanging information would not constitute a good approach to parallel
computations. Moreover, establishing a dedicated resource to be shared
between multiple processors, like shared memory in symmetric
multiprocessing, is generally not enough for an efficient parallel
operation. Additional means must be provided to facilitate interaction
among multiple processors. The following section covers how this
communication usually gets done in a multi-threaded environment.

Inter-Processor Communication and Multi-threaded Programming

The APIC plays a major role in communication among cores or
processors using the interprocessor interrupt (IPI). To illustrate
interprocessor communication for multi-threaded programming,
Figure 10.13 depicts how the IPI scheme is used in a parallel
programming environment.
One of the most important reasons an operating system uses IPIs is to
schedule the execution of threads on multiple processors. Normally, the
system may reschedule threads upon each timer or similar periodic
external interrupt. But, once a thread triggers a synchronization object
and this may happen at an unpredictable moment in time, not at all
periodicallyother threads waiting on that object need to be executed.
Of course, they may be queued and run at the next periodic interruption
signal, though in this case a considerable amount of time is wasted, from
the processors point of view, Another approach is to program the
threads in a manner that allows them to check the state of the other
threads they depend on, but such a programming scheme sometimes
cannot even be called parallel. Figure 10.14 illustrates these statements:
since all threads are executed in a preemptive environment, the model of
polling a state variable in memory does not always work wellthreads
may get preempted and the actual execution may be shifted in time with
regard to the moment when the state change occurred.
 Chapter 10: Threading on Intel Multi-Core Processors 267

Logical Logical Logical Logical Logical Logical Logical Logical

Processor 0 Processor 1 Processor 0 Processor 1 Processor 0 Processor 1 Processor 0 Processor 1

Processor Core Processor Core Processor Core Processor Core

Local APIC Local APIC Local APIC Local APIC Local APIC Local APIC Local APIC Local APIC

Bus Interface Bus Interface Bus Interface Bus Interface

Interrupts IPIs Interrupts IPIs Interrupts IPIs Interrupts IPIs

Processor System Bus

IPIs

Cluster IPIs, Interrupts

Bridge
Manager

PCI Bus

External Interrupts
I/O APIC

Chipset
IPIs

Processor System Bus

IPIs IPIs

Processor Processor

Figure 10.13 Relationship of Each Processors Local APIC, I/O APIC, and System Bus in a
Multiprocessor System

Processor 0 Processor 1 Processor N Processor 0 Processor 1 Processor N

Triggers a Polls the Executes Executes

Executes one another
semaphore semaphore Sends IPI
another application application
application
Executes Executes
worker worker Executes Executes Executes
Polls the
Time

thread 0 thread 1 worker worker worker

semaphore
thread 0 thread 1 thread N

Executes
worker
thread N

Figure 10.14 Use of Interprocessor Interrupts in Parallel Programming

268 Multi-Core Programming

Again, by employing the IPI scheme, the system ensures that all
waiting threads are executed immediately after the synchronization
object has been triggered, and executed with a predictable delay time,
much less than the normal rescheduling period.
But in some situations a threads wait time does not exceed the time
quantum granted to the thread by the operating system. In this case it
would be inefficient to reschedule the threads execution by returning
control to the OS or by making other threads issue an IPI to wake up
your waiting thread, since the interrupt delivery delay may be much
greater than the actual wait interval. The only solution would be to keep
control and wait for other threads in a loop. This is where the hardware
monitor/wait approach yields the best benefit to a programmer, because
one can boost the performance by providing a hint to the processor that
the current thread does not need any computational resources since all it
does is wait for a variable to change.

Power Consumption
You might be surprised to find this section in a software book. You
might know that you can control your system power by using available
system level APIs such as GetSystemPowerStatus and
GetDevicePowerState. Mobile application developers understand the
issue with power more than others. Traditionally, systems and
applications have been designed for high performance. In fact, our entire
discussion up to this point has been concerned with architectural and
programming innovations to increase performance of applications and
systems. However, recently the power consumption of the platform has
become a critical characteristic of a computing system.

Power Metrics
Increases in power consumption have occurred despite dramatic and
revolutionary improvements in process technology and circuit design. The
primary reason behind the increase in power has been the continued
emphasis on higher performance. As complexity and frequency of processors
has increased over the years to provide unprecedented levels of performance
the power required to supply these processors has increased steadily too. A
simplified equation that demonstrates power-performance relationship for
the CMOS circuits on which all modern processors are based is:
P ACV 2 f
 Chapter 10: Threading on Intel Multi-Core Processors 269

The first component in the equation captures the dynamic power of

charging and discharging transistor circuits from which basic functional
blocks of processors are built. The power is directly proportional to
switching frequency (f), square of the supply voltage (V) and total
capacitance (C). Since not all functional blocks are used at any given time
by a processor workload and not all gates are switched, A represents
activity factor or the number of switched transistors on a die.
The equation above demonstrates the trade-off between performance
and power. As processor frequency increases so does power consumption.
As processor architecture becomes more complex to support greater
levels of instruction level parallelism and increased performance,
capacitance of system increases and so does dynamic power.
Another and more subtle point is that concurrent processing can also
lead to significant power reduction. Splitting the workload into multiple
threads and running them in parallel can significantly increase processor
power efficiency. Multi-threading attacks two primary sources of energy
inefficiency that is related to activity factor A and total processor
capacitance C: unutilized resources and wasted resources due to
aggressive speculation in modern processors. Multi-threaded processors
rely much less on speculation and provide better resource utilization
leading to improved power efficiency.
The processors power efficiency can be measured and quantified in a
variety of ways. One common metric is peak power or Thermal Design
Power (TDP). This is the maximum power at which the processor can run
without exceeding thermal solution capabilities and damaging the part.
Another common metric used to quantify processor power consumption is
average power. It is usually computed as an average of instantaneous
power readings over execution of a certain benchmark. A more accurate
measurement of power efficiency is performance per watt. It may be
expressed in a variety of ways such as energy per instruction, MIPS/watts,
or benchmark performance score per average processor power
consumption. All these metrics reflect a fundamental interplay between
performance and power, and demonstrate the interdependency between
the two. The power efficiency metric is the fact that processors must be
efficient both in active and idle states. In reality, processors in mobile,
desktop, and server platforms spend a significant amount of time doing
nothing or being idle. Ensuring that processor power consumption is
minimal in this state is critical for the overall power efficiency. Note that a
processor with higher TDP or active power consumption will be more
energy efficient and potentially have lower average power if it provides a
very low power idle state. While it may be beneficial to expand power for
270 Multi-Core Programming

active workload for performance reasons, there is no benefit to wasting

power when you are doing nothing!

Reducing Power Consumption

The discussion so far has focused on the fundamental aspects and
importance of processor power consumption. It should be clear that
reducing processor power has become an overriding design goal. Achieving
low power processor operation is a complex task that requires effort at
multiple layers of hardware and software infrastructure. At the silicon level,
designers have developed advanced techniques such as strained silicon and
sleep transistors to reduce leakage and idle power. At the logic level,
extensive clock gating allows to turn off tree branches to latches and flip-
flops when they are not used saving considerable amount of power.
At the architecture level processors expose different frequency and
voltage settings that allow operating systems to adjust processor
performance to current workload demand. As the equation shows,
processor power is proportional to its frequency and square of the voltage.
Adjusting these two parameters to workload demand can lead to
significant power benefits without impacting perceived system
performance. In fact, frequency and voltage scaling is a feature in almost all

Intel processors today and is better known as Enhanced Intel SpeedStep
Technology. In addition, Intel processors support highly efficient
frequency and voltage transitions with rapid frequency scaling and no-stall
voltage changes. Since reaching low power state might not be worthwhile
if it takes a long time these capabilities are essential to maintaining system
performance and reducing dynamic power consumption.
When available, operating systems use Enhanced Intel SpeedStep
Technology to reduce platform power during operation. Both Linux and
Windows implement similar algorithms. Operating systems determine the
right frequency setting by measuring the time they spend in idle loop.
Whenever the processor is underutilized the OS determines the
frequency that will increase processor utilization to a certain level. On
the other hand, if processor utilization becomes higher than a certain
threshold, the OS will increase processor frequency to meet higher
demand. In todays operating system, power management is governed by
Advance Configuration and Power Interface (ACPI). ACPI is part of BIOS
and exposes power management features and details to the operating
system. All Intel processors today provide support for a variety of idle
and performance states enabling operating systems to take advantage of
these features for power efficient operation.
 Chapter 10: Threading on Intel Multi-Core Processors 271

Finally, in addition to processors and operating systems, applications

have a responsibility to reduce power consumption of computing devices.
With the proliferation of multi-core processors, application developers
have a great opportunity to decrease system power consumption by
writing efficient and optimized multi-threaded applications. Multi-core
processors offer clear performance benefits with close to linear speedup as
you add cores. With the performance advantage coming from parallelism
and not from an increase in frequency and voltage, multi-core processors
can be very energy-efficient. By writing multi-threaded applications that
take full advantage of parallelism provided by the cores, developers can
cause processor throughput to be dramatically increased, more than
making up for greater power dissipation due to a larger number of cores,
and resulting in significant net gain in power efficiency.

Beyond Multi-Core Processor Architecture

Every technology has a next step. Things go from conception to research
and move on to development. The progression continues to evolve. You
now know that having more than a single core is referred to as multi-core. In
theory, the number of cores is only limited by the availability of supporting
technologies. Over time we can anticipate processors with more and more
cores as we continue to innovate. Intel is working to deliver innovative
processor technology on a roadmap as proposed in Figure 10.15.

10 10000
Nominal feature size
0.7X every
2 years

1 1000

Micron 130nm Nano-

90nm meter
Gate Length 65nm
0.1 45nm 100
70nm 32nm
22nm
Nanotechnology 50nm
(<100nm) 35nm
25nm
18nm
0.01 12nm 10
1970 1980 1990 2000 2010 2020
Future options subject to change

Figure 10.15 Progression of Silicon Technology towards Nanoscale

272 Multi-Core Programming

The silicon technology is also approaching the nanoscale domain to

enhance transistor density as well, as shown in Figure 10.16.

90 nm
2003 65 nm Technology Generation
Manufa
cturing 2005 45 nm
20 32 nm
Develo 07 2009
pment
2011+
Resea
rch

SiGe S/D
Strained SiGe S/D
Silicon Strained
Silicon

Future options subject to change

Figure 10.16 Innovation-enabled Technology Pipeline

Beyond nanoscale, the technology proposes to drive transistors to the

quantum level. If the progress continues, you can expect processors with
many cores before the end of this decade. As the number of cores
increases, the processor architecture will provide significant hardware-
based, thread-level, parallel capability on a single processor. With these
levels of compaction and multiplicity, just imagine the possibilities for
the future of multi-core processors.

Key Points
When developing a software application, the focus usually remains on
the implementation layer and the layer below. Several layers separate the
abstracted application and the hardware. With the recent development of
more than one core in a single package, developers have to consider
every component in the solution domain to optimize the capabilities of
these new processors.
 Chapter 10: Threading on Intel Multi-Core Processors 273

Key concepts to remember about threading on multi-core processors:

There are four types of threading models used in a processor:
fine-grained, coarse-grained, SMT, and CMT.
For the fine-grained threading model, context switching occurs in
every cycle and for the coarse-grained threading model, context
switching occurs when a pipeline gets stalled. On the other hand,
there is no context switching required for true SMP or multi-core
processors.
Processors have enough resources to handle a good number of
instructions in flight during full operation.
Intel has two types of general processor architecture: wide
speculative superscalar and EPIC. Superscalar processors are used
for Intel 32-bit processors and Intel EM64T. EPIC is used for
Itanium processors.
Recently, an implementation of HT Technology was introduced
in the Itanium processor as well.
Interprocessor or intercore communication is done by inter-
processor interrupt (IPI) with the help of Advanced
Programmable Interrupt Controller (APIC) features in the
processors.
Advance Configuration and Power Interface (ACPI) exposes
processor frequency and voltage levels as performance states.
The power factors can be controlled through software APIs.
Enhanced Intel SpeedStep Technology is available on platforms to
improve power-based performance.
 Chapter 11

Intel Software
Development
Products
W riting a threaded application requires the same create, debug, and
tuning steps needed to make a working application that is not
threaded. While so much is the same, it is most interesting to look at
what is different.
This chapter takes a look at Intels suite of software and focuses on
the aspects of these products that Intel has included for threaded
applications. Intels suite of products is arguably the most comprehensive
available today for threaded programming for C++ or Fortran developers,
including a few tools that are currently unique or leading examples.
Most of the tools are focused on threading. A section in this chapter
also describes Message Passing Interface (MPI) programming and the
tools to support it. MPI programming is an important programming
method to consider when you are trying to make highly scalable code
that might even run on a very large supercomputer.

Overview
Intel has been working with multiprocessor designs and the tools to
support them for well over a decade. In order to assist programmers,
Intel has made available a number of tools for creating, debugging, and
tuning parallel programs.

275
276 Multi-Core Programming

Investigate
Most programming work begins with an existing application. It often
begins with a prototype of a framework or critical elements of the
application, for those working to program something entirely new.
Whether a prototype or a preexisting application, some initial
investigation plays a critical role in guiding future work. Tools such

as the Intel VTune Performance Analyzer and the Intel Thread
Profiler are extremely useful. The Intel compilers can play a strong
role in what if experiments by simply throwing some switches, or
inserting a few directives in the code, and doing a recompile to see
what happens.

Create/Express
Applications are written in a programming language, so a compiler is a
natural place to help exploit parallelism. No programming languages
in wide usage were designed specifically with parallelism in mind.
This creates challenges for the compiler writer to automatically find
and exploit parallelism. The Intel compilers do a great deal to find
parallelism automatically. Despite this great technology, there are too
many limitations in widely used programming languages, and
limitations in the way code has been written for decades, for this
to create a lot of success. Automatic parallelization by the compiler
is nevertheless a cheap and easy way to get some help
all automatically. Auto-parallelization is limited by all popular
programming languages because the languages were designed without
regard to expressing parallelism. This is why extensions like OpenMP
are needed, but they are still limited by the programming languages
they extend. There is no cheap and easy way to achieve parallelism
using these languages.
To overcome limitations imposed by conventional programming
languages, the Intel compilers support OpenMP, which allows a
developer to add directives to the code base that specify how different
code segments may be parallelized. This allows programs to get
significant performance gains in a simple, easy-to-maintain fashion. The
OpenMP extensions have been covered in some detail in Chapter 6. Intel
libraries also help make the production of threaded applications easier. In
this case, Intel engineers have done the work for you and buried it in the
implementation of the libraries. These may be the very same libraries you
were using before threading.
 Chapter 11: Intel Software Development Products 277

Debugging
Having multiple threads combine to get the work of an application done
gives rise to new types of programming errors usually not possible with
single threaded applications. Up until recently, these threading errors
were simply bugs that needed to be debugged the old fashion wayseek
and find. With the Intel Thread Checker, developers can directly locate
threading errors. It can detect the potential for these errors even if the
error does not occur during an analysis session. This is because a well-
behaved threaded application needs to coordinate the sharing of memory
between threads in order to avoid race conditions and deadlock. The
Intel Thread Checker is able to locate examples of poor behavior that
should be removed by the programmer to create a stable threaded
application.

Tuning
Performance tuning of any application is best done with non-intrusive
tools that supply an accurate picture of what is actually happening on a
system. Threaded applications are no exception to this. A programmer,
armed with an accurate picture of what is happening, is able to locate
suboptimal behavior and opportunities for improvement. The Intel
Thread Profiler and the Intel VTune Performance Analyzer help tune a
threaded application by making it easy to see and probe the activities of
all threads on a system.

Intel Thread Checker
The Intel Thread Checker is all about checking to see that a threaded
program is not plagued by coding errors in how threads interoperate that
can cause the program to fail. It is an outstanding debugging tool, even
for programs that seem to be functioning properly. Just knowing that
such a tool exists is a big step since this is such a new area for most
programmers. Finding this class of programming error is especially
difficult and frustrating because the errors manifest themselves as
nondeterministic failures that often change from run to run of a program
and most often change behavior when being examined using a debugger.
Developers use the Intel Thread Checker to locate a special class of
threading coding errors in multi-threaded programs that may or may not
be causing the program to fail. The Intel Thread Checker creates
diagnostic messages for places in a program where its behavior in a
278 Multi-Core Programming

multi-threaded environment is potentially nondeterministic. The Intel

Thread Checker identifies issues including data races, deadlocks, stalled
threads, lost signals and abandoned locks. The Intel Thread Checker
supports analysis of threaded programs that use OpenMP, POSIX, and the
Windows API.
Chapter 7 explained data races and deadlocks which are the two
programming errors that can occur because of threading. These are difficult
to debug as they can cause results to be indeterminate and to differ from
the output that a non-threaded version of the program would produce.
Deadlock causes a program, or a subset of threads, to not be able to
continue executing at all because of errors in the way it was programmed.
The process of finding critical multi-threading programming issues
like data races and deadlocks starts with running a program with the Intel
Thread Checker to collect data. Once this data collection has occurred,
the Intel Thread Checker is used to view the results of the program
execution. These results are shown in a prioritized list of diagnostic and
warning messages based on the trace data. Sorting and organizing the
Diagnostics list in various ways helps focus on the most important issues.
This tool isolates threading bugs to the source code line where the bug
occurs. It shows exactly where in a program threading errors are likely to
happen. When the Intel Thread Checker detects an issue, it reports the
function, context, line, variable, and call stack to aid in analysis and
repair. It also provides a suggestion of possible causes for the threading
errors and suggested solutions with one-click diagnostic help.
The Intel Thread Checker suggests all necessary warnings for effective
threaded application diagnosis, while allowing you to choose which
warnings to display at different points in the product development cycle.

How It Works
The Intel Thread Checker can do its analysis using built-in binary
instrumentation and therefore can be used regardless of which compiler is
used. This is particularly important with modern applications that rely on
dynamically linked libraries (DLLs) for which the source code is often
unavailable. The Intel Thread Checker is able to instrument an application
and the shared libraries, such as DLLs, that the application utilizes.
When combined with the Intel compiler and its compiler-inserted
instrumentation functionality, Intel Thread Checker gives an even better
understanding by making it possible to drill down to specific variables on
each line. Figure 11.1 shows the diagnostic view and Figure 11.2 shows
the source view of the Intel Thread Checker.
 Chapter 11: Intel Software Development Products 279

Figure 11.1 Intel Thread Checker Diagnostic View

Figure 11.2 Intel Thread Checker Source View

280 Multi-Core Programming

Usage Tips
Because the Intel Thread Checker relies on instrumentation, a
program under test will run slower than it does without
instrumentation due to the amount of data being collected.
Therefore, the most important usage tip is to find the smallest data
set that will thoroughly exercise the program under analysis.
Selecting an appropriate data set, one that is representative of your
code without extra information, is critical so as not to slow execution
unnecessarily. It is generally not practical to analyze a long program
or run an extensive test suite using this tool.
In practice, three iterations of a loopfirst, middle, and lastare
usually sufficient to uncover all the problems that the Intel Thread
Checker is able to find within each loop. The exception is when if
conditions within the loop do different things for specific iterations.
Because of the overhead involved in the Intel Thread Checker operation,
you should choose a data set for testing purposes that operates all the
loops that you are trying to make run in parallel, but has the smallest
amount of data possible so that the parallel loops are only executed a
small number of iterations. Extra iterations only serve to increase the
execution time. If you have a particular section of code you would like to
focus on, you can either craft your data and test case to exercise just that
part, or you can use the Pause/Resume capabilities of the Intel Thread
Checker.
The Intel Thread Checker prioritizes each issue it sees as an error,
warning, caution, information, or remark, as shown in Figure 11.3.
Sorting errors by severity and then focusing on the most important issues
first is the best way to use the tool.
Before you prepare your code for use with the Intel Thread
Checker, you should ensure that your code is safe for parallel
execution by verifying that it is sequentially correct. That is, debug it
sequentially before trying to run in parallel. Also, if your language or
compiler needs special switches to produce thread-safe code, use
them. This comes up in the context of languages like Fortran, where
use of stack (automatic) variables is usually necessary, and not always
the default for a compiler. The appropriate switch on the Intel Fortran
Compiler is Qauto. Use of this option on older code may cause
issues, and the use of a SAVE statement in select places may be
required for subroutines that expect variables to be persistent from
invocation to invocation.
 Chapter 11: Intel Software Development Products 281

Severity

Diagnostic Groups
2
Unclassified
Remark
Information
Caution
Warning
Error
Filtered

0 1 2 3 4 5
Number of Occurences

Figure 11.3 Intel Thread Checker Bar Chart with Error Categories

Using Intel Thread Checker with OpenMP

OpenMP programs are threaded programs and can suffer from the same
errors and performance problems as explicitly threaded applications.
OpenMP is discussed in detail in Chapter 6. Using Intel Thread Checker,
you can avoid the standard task of identifying storage conflicts in
previously threaded code. With OpenMP, the diagnostic output of Intel
Thread Checker identifies and allows the categorizing of the scope of
variables within parallel regions. This allows a programmer using
OpenMP to try a directive that is close to correct and fine tune it using
Intel Thread Checker. Intel has provided a whitepaper on this exact topic

on their developer web site, www.intel.com/software, titled Intel
Threading Tools and OpenMP by Clay P. Breshears.

Intel Compilers
Just as with previous hardware technologies, the compiler can
play a central or supporting role in taking advantages the
multi-processor/multi-core/multi-threading capabilities of your shared
282 Multi-Core Programming

memory platform. Intel Corporation has been supporting multi-

threading capabilities in its compilers for many years now. This section
explores those technologies.

OpenMP
In Chapter 6, you learned how OpenMP can be used as a portable
parallel solution and that Intel compilers have support for OpenMP
within the Windows and Linux environments. Intel compilers support
all the implementation methodologies discussed in Chapter 6. At the
time of this writing, Version 9.1 of the Intel compilers support the
OpenMP API 2.5 specification as well as the workqueuing extension, a
feature proposed by Intel for OpenMP 3.0. To get the Intel compiler to
recognize your OpenMP constructs, compile with the following
switch:
Windows: /Qopenmp
Linux: -openmp

Some of the many advantages of using OpenMP to thread software

are:
It is intuitive and comparatively easy to introduce into your
application.
It is portable across operating systems, architectures, and
compilers.
The compiler has the opportunity to make architecture-specific
optimizations.
OpenMP API achieves these goals by leaving the implementation up to
the compiler.

Atomic
The OpenMP Atomic directive is probably the most obvious example of a
feature where the compiler is able to provide a fast implementation. The
atomic directive is similar to a critical sectionin that only one thread
may enter the atomic section of code at a timebut it places the
limitation on the developer that only very simplistic and specific
statements can follow it.
When you use the atomic directive as follows:
#pragma omp atomic
workunitdone++;
 Chapter 11: Intel Software Development Products 283

The compiler can issue the following instructions that allow the
hardware to atomically add one to the variable
mov eax, 0x1h
lock xadd DWORD PTR [rcx], eax

This is much more efficient than locking the code using a critical section
or a mutex, then updating the variable, and finally releasing the lock,
which can take hundreds or thousands of cycles, depending on the
implementation. This could be created using inline assembly or compiler
intrinsics, except that then the code would not be portable to other
architectures or OS environments.
The Intel compilers will perform other optimization algorithms when
compiling OpenMP code. The atomic example was chosen due to its
simplicity. As optimization techniques are developed by Intels compiler
developers, those techniques usually get added in the compiler so that
everyone who uses OpenMP with the Intel compiler benefits, whether
they are aware of it or not.

Auto-Parallel
The Intel compilers have another feature to help facilitate threading. The
auto-parallelization feature automatically translates serial source code into
equivalent multi-threaded code. The resulting binary behaves as if the
user inserted OpenMP pragmas around various loops within their code.
The switch to do this follows:
Windows: /Qparallel
Linux: -parallel

For some programs this can yield a free performance gain on SMP
systems. For many programs the resulting performance is less than
expected, but dont give up on the technology immediately. There are
several things that can be done to increase the probability of
performance for this auto-parallel switch.
Increasing or decreasing the threshold for which loops will be made
parallel might guide the compiler in creating a more successful binary.
The following switch guides the compiler heuristics for loops:
Windows: /Qpar_threshold[:n]
Linux: -par_threshold[n]

where the condition 0 <= n <= 100 holds and represents the
threshold for the auto-parallelization of loops. If n=0, then loops get
284 Multi-Core Programming

auto-parallelized always, regardless of computation work volume. If

n=100, then loops get auto-parallelized when performance gains are
predicted based on the compiler analysis data. Loops get auto-parallelized
only if profitable parallel execution is almost certain. The intermediate
values 1 through 99 represent the percentage probability for profitable
speed-up. For example, n=50 directs the compiler to parallelize only if
there is a 50 percent probability of the code speeding up if executed in
parallel.
Using auto-parallelization in combination with other switches like
Inter-Procedural Optimizations (IPO), Profile Guided Optimizations
(PGO) and High Level Optimizations (HLO) aids the compiler in making
more correct choices while threading the code.
If auto-parallelization does not help directly, it can perhaps help
indirectly. The Intel compilers also support a compiler reporting feature
and the switch is:
Windows: /Qpar_report[n]
Linux: -par_report[n]

where 0 <= n <= 3. If n=3, then the report gives diagnostic information
about the loops it analyzed. The following demonstrates the use of this
report on a simplistic example. Given the following source:
1 #define NUM 1024
2 #define NUMIJK 1024
3 void multiply_d( double a[][NUM], double b[][NUM],
4 double c[][NUM] )
5 {
6 int i,j,k;
7 double temp;
8 for(i=0; i<NUMIJK; i++) {
9 for(j=0; j<NUMIJK; j++) {
10 for(k=0; k<NUMIJK; k++) {
11 c[i][j] = c[i][j] + a[i][k] * b[k][j];
12 }
13 }
14 }
15 }
The compiler produces the following report:
$ icc multiply_d.c -c -parallel -par_report3
procedure: multiply_d
serial loop: line 10: not a parallel candidate due to insufficent
work
serial loop: line 8
anti data dependence assumed from line 11 to line 11, due to "b"
anti data dependence assumed from line 11 to line 11, due to "a"
 Chapter 11: Intel Software Development Products 285

flow data dependence assumed from line 11 to line 11, due to "c"
flow data dependence assumed from line 11 to line 11, due to "c"
serial loop: line 9
anti data dependence assumed from line 11 to line 11, due to "b"
anti data dependence assumed from line 11 to line 11, due to "a"
flow data dependence assumed from line 11 to line 11, due to "c"
flow data dependence assumed from line 11 to line 11, due to "c"

Based on this report, you can see the compiler thinks a dependency
exists between iterations of the loop on the a, b, and c arrays. This
dependency is due to an aliasing possibilitybasically, it is possible that
the a or b array points to a memory location within the c array. It is easy
1
to notify the compiler that this is not possible . To handle such instances,
any of the following techniques can be used:
Inter-Procedural Optimization (IPO)
Windows: /Qipo
Linux: -ipo
Restrict keyword
Aliasing switches: /Oa, /Ow, /Qansi_alias
#pragma ivdep
After modifying the code as follows:

void multiply_d(double a[][NUM], double b[][NUM], double c[restrict][NUM])

the following report is produced:

$ icc multiply_d.c -c -parallel -par_report3 -c99
procedure: multiply_d
serial loop: line 10: not a parallel candidate due to insufficent work
multiply_d.c(8) : (col. 2) remark: LOOP WAS AUTO-PARALLELIZED.
parallel loop: line 8
shared : { "c" "b" "a" }
private : { "i" "j" "k" }
first priv.: { }
reductions : { }

1
In this case, the programmer assumes the responsibility of ensuring that this aliasing doesnt occur.
If the programmer is wrong, unpredictable results will occur.
286 Multi-Core Programming

This technique can also be used as a guide in adding OpenMP pragmas to

the source. For the above example, the following OpenMP changes are
easy to identify:

1 #define NUM 1024

2 #define NUMIJK 1024
3 void multiply_d( double a[][NUM], double b[][NUM],
4 double c[][NUM] )
5 {
6 int i,j,k;
7 double temp;
8 #pragma omp parallel for shared(a,b,c) private(i,j,k)
9 for(i=0; i<NUMIJK; i++) {
10 for(j=0; j<NUMIJK; j++) {
11 for(k=0; k<NUMIJK; k++) {
12 c[i][j] = c[i][j] + a[i][k] * b[k][j];
13 }
14 }
15 }
16 }
The auto-parallelization feature of the Intel compilers may provide an
easy performance gain in your source. If it doesnt, you can increase its
probability of helping you guide it with other switches, aliasing
techniques, or by using it to guide the insertion of OpenMP pragmas. If
a specific portion of the application does not thread through
auto-parallelizationor if it does thread the code, but it does so
inefficientlyreport this to Intel through Intel Premier Support Web site.
It is possible that the compiler developers can add that optimization to
the compiler, thereby making the application run faster, as well as
improving the compiler for the overall community.

Software-based Speculative Precomputation

Version 9.0 of the Intel compilers introduced a preview feature called
Software-based Speculative Precomputation (SSP), also known as Helper
Threads. The goal of SSP is to hide memory latencies associated with
single-threaded applications by utilizing idle or unused multi-threading
hardware resources to prefetch data from memory into the cache. In
order to do the prefetch, the compiler creates secondary thread(s) that
run on behalf of the main thread. The secondary thread or threads try to
access data in memory that will soon be needed by the main thread. If
the needed data is not currently in the cache, a cache miss occurs and
the data is loaded into the cache. In the ideal case, the data will be in the
 Chapter 11: Intel Software Development Products 287

cache before the main thread needs the data. Since the hardware
threading resources would have been idle otherwise, this technique
effectively eliminates performance penalties associated with memory
latencies. This technique will work for any system that can execute
threads simultaneously and includes a shared cache that multiple threads
can access directly.
In order for this technique to yield a performance gain the compiler
needs detailed data about cache misses within your application. The
compiler needs to gather an execution profile of your application and
data on cache misses from the Performance Monitoring Unit (PMU) in
order to identify where cache misses are occurring in your application.

Compiler Optimization and Cache Optimization

In order to achieve the maximum benefit from threading, it is also
important to make sure your application is optimized for the underlying
hardware platform. Two aspects that are relevant to threading should be
considered:
Increasing cache usage (thereby decreasing bus bandwidth)
Increasing the performance of every thread
One of the performance-limiting factors on a parallel processing
capable system is the memory subsystem bottleneck. The Intel
compiler can help avoid main memory accesses by performing
optimizations within the compiler to maximize the use of the caches
which ultimately decreases the amount of data that needs to pass
through the memory bus.
On some architectures it is often optimal to reduce the use of
prefetching, as this can cause unnecessary accesses to memory. When
the Intel compiler uses OpenMP or Auto-Parallelization, the compiler may
reduce using prefetches on architectures where this is relevant.
The default switches in the Intel compiler may not yield the optimal
performance or the best cache optimizations. The compiler has several
features that can increase the probability that your application will
perform better:
Higher Optimization Levels (/O1,/O2,/O3)
Vectorization (/Q[a]xP, /Q[a]xN, Q[a]xW, /Q[a]xB, /Q[a]xK)
Inter-Procedural Optimizations (/Qipo)
Profile Guided Optimizations (/Qprof_gen -> /Qprof_use)
288 Multi-Core Programming

Intel Debugger
Chapter 8 covered a number of general purpose debugging techniques
for multi-threaded applications. In order to provide additional help to
developers Intel has developed a debugging tool appropriately named the
Intel Debugger (IDB). The Intel Debugger is shipped as part of the Intel
compilers. It is a full-featured symbolic source-code application debugger
that helps programmers to locate software defects that result in run-time
errors in their code. It provides extensive debugging support for C, C++
and Fortran, including Fortran 90. It also provides a choice of control
from the command line, including both dbx and gdb modes, or from a
graphical user interface, including a built-in GUI, ddd, Eclipse CDT, and
Allinea DDT.
The Intel compilers enable effective debugging on the platforms they
support. Intel compilers are debugger-agnostic and work well with
native debuggers, the Intel Debugger, and selected third-party debuggers.
By the same token, the Intel Debugger is compiler-agnostic and works
well with native compilers, the Intel compilers, and selected third-party
compilers. This results in a great deal of flexibility when it comes to
mixing and matching development tools to suite a specific environment.
In addition, the Intel Debugger provides excellent support for the
latest Intel processors, robust performance, superior language-feature
support, including C++ templates, user-defined operators, and modern
Fortran dialects (with Fortran module support); and support for Intel
Compiler features not yet thoroughly supported by other debuggers.
The Intel Debugger is a comprehensive tool in general and also
supports extensively for threaded applications as well. Some of the
advanced capabilities of the Intel Debugger for threaded applications are:
Includes native threads and OpenMP threads
Provides an all threads stop / all threads go execution model
Acquires thread control on attach and at thread creation
Ability to list all threads and show indication of thread currently
in focus
Set focus to a specific thread
Sets breakpoints and watchpoints for all threads or for a subset of
all threads (including a specific thread)
Most commands apply to thread currently in focus or to any/all
threads as appropriate
 Chapter 11: Intel Software Development Products 289

Optional thread-specific qualifier for many commands

Access to Thread Local Storage and Shared Local Variables
For the cluster-parallel applications, the Intel Debugger makes use of MPI
including a proprietary cluster aggregation network and support for user-
defined process sets that can be stopped or moved forward
independently of one another.
Even though the Intel Debugger does not support examining the
content of mutexes and condition variables, it can be used from the
command line to call directly into native thread libraries and OpenMP
libraries for more detailed information.

Intel Libraries
Libraries are an ideal way to utilize parallelism. The library writer can hide
all the parallelism and the programmer can call the routines without
needing to write parallel code. Intel has two libraries that implement
functions that have been popular for many years, and which Intel has
gradually made more and more parallel leading up to today when they are
parallelized to a great extent. Both of Intels libraries are programmed using
OpenMP for their threading, and are pre-built with the Intel compilers. This
is a great testimonial to the power of OpenMP, since these libraries produce
exceptional performance using this important programming method.

Intel Math Kernel Library

The Intel Math Kernel Library (Intel MKL) is a set of highly optimized
routines used for mathematical problem solving. The routine are used for
solving problems of computational linear algebra, performing the
discrete Fourier transforms, and solving some other computation-
intensive problems. The library includes routines of the BLAS, Sparse
BLAS, LAPACK and ScaLAPACK packages (Fortran interfaces), sparse
solver, interval linear solvers, CBLAS (C interface to BLAS routines), as
well as discrete Fourier (with Cluster DFTI) and fast Fourier transform
routines, vector mathematical functions, and the Vector Statistical Library
(Fortran and C interfaces for random number generators and
convolution/correlation mathematical operations). The library functions
ensure high performance when run on Intel processors or compatible
processors. Level 3 BLAS and most LAPACK routines, in particular, take
advantage of multiprocessor computation through threading.
290 Multi-Core Programming

Intel MKL is threaded in a number of places: sparse solver, LAPACK

(*GETRF, *POTRF, *GBTRF, *GEQRF, *ORMQR, *STEQR, *BDSQR
routines), all Level 3 BLAS, Sparse BLAS matrix-vector and matrix-matrix
multiply routines for the compressed sparse row and diagonal formats,
and all discrete Fourier transform (DFT) routinesexcept 1D
transformations when DFTI_NUMBER_OF_TRANSFORMS=1 and sizes are
not a power-of-two, and all fast Fourier transform (FFT) routines.

Intel Integrated Performance Primitives
The Intel Integrated Performance Primitives (Intel IPP) are a set of highly
optimized routines used as the basis for much multimedia
encode/decode work as well as a variety of other nonscientific problems.
The routines are used for image processing, audio coding, speech coding,
JPEG, video coding, speech recognition, color conversion, computer
vision, data compression, signal processing, cryptography, string
processing, matrix processing and vector math. The library functions
ensure high performance when run on Intel processors or compatible
processors. Higher level routines take advantage of multiprocessor
computation through threading. Each release has a list of functions that
take advantage of threading; the IPP 5.0 release, for instance, lists 563
routines that use OpenMP to offer parallelism.

Parallel Program Issues When Using Parallel Libraries

Using a parallel library from a program that expresses some parallelism
itself creates a situation where some thought is required. Running a four-
processor machine is most efficient if a program uses four active threads.
However, if the program is actively running four threads in a parallel
region, and each thread calls a library routine that in turn tries to spawn
four threads, the elegance disappears and conflict arises. Therefore, it is
critical that a developer understand not only the parallel nature of their
application, but the underlying parallel implementation of any external
libraries used in implementing that application.
If the user threads the program using OpenMP directives and uses
the Intel compilers to compile the program, Intel Math Kernel Library
(Intel MKL) and the user program will both use the same threading
library. This solves many potential issues automatically for the user.
Intels libraries will determine if the function is called while in a parallel
region in the program, and if it is, it does not spread its operations over
multiple threads. However, the Intel libraries can be aware that the
 Chapter 11: Intel Software Development Products 291

function is in a parallel region only if the threaded program and the

library are using the same threading library. If the user program is
threaded by some other means, the library may operate in multi-
threaded mode and the computations may be slow or possibly
corrupted. This implies that the programmer should take the following
into consideration:
If the user threads the program using OS threads (pthreads on
Linux, Win32 threads on Windows), and if more than one thread
calls the library, and the function being called is threaded, it is
important that threading in Intel MKL be turned off. Set
OMP_NUM_THREADS=1 in the environment. This is the default
with Intel MKL except for sparse solver.
If the user threads the program using OpenMP and compiles the
program using a compiler other than a compiler from Intel, then
the best approach is to force the library to run in serial. This case
is more problematic than the previous one in that setting
OMP_NUM_THREADS in the environment affects both the
compiler's threading library and the threading library used by the
Intel libraries. For Intels Math Kernel Library, you set
MKL_SERIAL=YES, which forces Intel MKL to serial mode
regardless of OMP_NUM_THREADS value.
If multiple programs are running on a multiple-CPU system, as in
the case of a parallelized program running using MPI for
communication in which each processor is treated as a node,
then the threading software will see multiple processors on the
system even though each processor has a separate process
running on it. In this case, OMP_NUM_THREADS should be set
to 1 to force serial use of the libraries and defer to the wisdom
of the programmer to orchestra the system usage at a higher
level.

The Future
Libraries will expand as a popular method for achieving parallelism. The
need for more standardizationfor compilers, users, and libraries to
cooperate with regards to the creation and activation of threadswill
grow. Right now, a careful programmer can pour through the
documentation for libraries and compilers and sort out how to resolve
potential conflicts. As time passes, we hope to see some consensus on
how to solve this problem and make programming a little easier. We will
292 Multi-Core Programming

see the emergence of more domain specific parallel libraries, as well as

some general frameworks built around libraries.

Intel Threading Building Blocks

Intel is developing a new approach to help developers with libraries. The
Intel Threading Building Blocks are a higher-level abstraction for threaded
applications that will also be understandable by analysis tools. Since the
project is currently in development, consult the Intel Web site for more
information on features and availability.

Intel VTune Performance Analyzer
The Intel VTune Performance Analyzer is a system-wide analysis tool that
offers event sampling and call graphs that include all information
available broken down not only by processes/tasks, but also by the
threads running within the processes. Intel Press offers a whole book on
the Analyzer, which dives into its numerous capabilities. This section
gives you just a flavor for the features, and highlights some of the ways
the Analyzer feature can be used in the tuning of threaded applications.
Users have summed up the tool by saying that the VTune analyzer
finds things in unexpected places. Users of the VTune analyzer are
enthusiastic about this tool largely because of this remarkable capability.
Threading adds a dimension to already complex modern computer
systems. It is no surprise when things happen on a system that cannot be
easily anticipated. When you seek to refine a computer system, the best
place to start is with a tool that can find these hidden problems by giving
a comprehensive performance exam.
Measurements are the key to refinement. The Intel VTune
Performance Analyzer is a tool to make measurements. It also has
wonderful features to help you understand those measurements, and
even advises you on what exceptional values may mean and what you
can do about them.
Taking a close look at the execution characteristics of an application
can guide decisions in terms of how to thread an application. Starting
with the hotspotsthe main performance bottlenecksin the
application, one can see if threading can be applied to that section of
code. Hotspots are found using the event sampling features in the VTune
analyzer. If the hotspot is in a location with little opportunity for
parallelism, a hunt up the calling sequence will likely find better
 Chapter 11: Intel Software Development Products 293

opportunities. The calling sequence can be traced back using the call-
graph capability of the analyzer. Implementations of threads can be
refined by looking at balance using the Samples Over Time feature and
the Intel Thread Profiler in the analyzer.

Find the Hotspot

The VTune analyzer can find the modules, functions, threads and even
the line of source code that consume most of the CPU cycles without
requiring a special build of the application. Source code displays require
a version with symbol information not stripped outthe default on
Linux, and needs a special option on Windows.
Shown here in Figure 11.4 is the module view in the analyzer for
Windows after collecting sampling data on the platform. It shows the
majority of CPU being spent in our program. Additional mouse clicks will
reveal functions and even the source code lines.

Figure 11.4 Sampling Results Using the Intel VTune Performance Analyzer
294 Multi-Core Programming

If you can distribute the work currently done on one processor onto
two processors, you can theoretically double the performance of an
application. Amdahls law reminds us that we cannot make a program
run faster than the sequentialnot written to run in parallelportion of
the application, so dont expect to leap to doubled performance every
time.

Using Call Graph for Finding a Threading Point

The VTune analyzer includes a Call Graph feature to create a call
graph of an application. By looking at the call graph as shown in
Figure 11.5, you can find places farther up in the call tree from the
hotspot in a function where it could make sense to create a thread. By
rewriting a higher level location in a program to partition the work
among several threads, parallel processing should improve the
performance of the application.

Figure 11.5 Call Graph Results, Viewed by Thread

 Chapter 11: Intel Software Development Products 295

Check the Load Balancing

How well distributed and parallelized the workload is can be examined
using the VTune analyzers Samples Over Time feature. The Samples
Over Time display shows how, for a particular application, module, or
thread, the time data was collected, as shown in Figure 11.6. Looking at
the thread sampling data over time shows if the CPU time consumed by
each thread was about the same, providing evidence of whether the
workload was evenly distributed. You can also look to see if the number
of samples taken by each thread was significant and in the same range.

Figure 11.6 Example of a Sampling Over Time View

Intel Thread Profiler
The Intel Thread Profiler is implemented as a view within the VTune
Performance Analyzer, but it is so significant that it should be discussed
as if it were an entirely separate product. Unlike the rest of the VTune
analyzer, the Intel Thread Profiler is aware of synchronization objects
used to coordinate threads. Coordination can require that a thread wait,
so knowing about the synchronization objects allows Intel Thread
Profiler to display information about wait time, or wasted time. The Intel
Thread Profiler helps a developer tune for optimal performance by
providing insights into synchronization objects and thread workload
imbalances that cause delays along the longest flows of execution.
The Intel Thread Profiler shows an applications critical path as it
moves from thread to thread, helping a developer decide how to use
threads more efficiently, shown in Figure 11.7. It is able to identify
synchronization issues and excessive blocking time that cause delays for
Win32, POSIX threaded and OpenMP code. It can show thread workload
imbalances so a developer can work to maximizes threaded application
performance by maximizing application time spent in parallel regions
296 Multi-Core Programming

doing real work. Intel Thread Profiler has special knowledge of OpenMP,
and can graphically display the performance results of a parallel
application that has been instrumented with calls to the OpenMP
statistics-gathering run-time engine.

Figure 11.7 The Intel Thread Profiler Critical PathTimeline View

The Timeline view shows the contribution of each thread to the total
program, whether on the critical path or not. The Thread Profilers also
has the ability to zero in on the critical path: the Critical Paths view
shows how time was spent on your program's critical path, the Profile
view displays a high-level summary of the time spent on the critical path.
Using the VTune Performance Analyzer and the Intel Thread Profiler
together, provide insights for a developer about threading in their
applications and on their systems. Together, these analysis tools help the
developer avoid searching for opportunities through trial and error by
providing direct feedback.

MPI Programming
Threading is a convenient model where each thread has access to the
memory of the other thread. This is portable only between shared
memory machines. In general, parallel machines may not share memory
between processors. While this is not the case with multi-core
processors, it is important to point out that parallel programs need not be
written assuming shared memory.
When shared memory is not assumed, the parts of a program
communicate by passing messages back and forth. It is not important
 Chapter 11: Intel Software Development Products 297

how the messages are passed; the details of the interconnect are hidden
in a library. On a shared memory machine, such as a multi-core
processor, this is done through shared memory. On a supercomputer
with thousands of processors, it may be done through an expensive and
very high speed special network. On other machines, it may be done via
the local area network or even a wide area network.
In order for a message-passing program to be portable, a standard for
the message passing library was needed. This formed the motivation
behind the Message Passing Interface (MPI), which is the widely used
standard for message passing. Many implementations exist including
vendor-specific versions for their machines or interconnects. The two
most widely used versions of MPI are MPICH, with roots from the earliest
days of UNIX and now hosted by Argonne National Lab, and LAM/MPI, an
open-source implementation hosted by Indiana University.
MPI makes possible source-code portability of message-passing
programs written in C, C++, or Fortran. This has many benefits, including
protecting investments in a program, and allowing development of the
code on one machine such as a desktop computer, before running it on
the target machine, which might be an expensive supercomputer with
limited availability.
MPI enables developers to create portable and efficient programs
using tightly coupled algorithms that require nodes to communicate
during the course of a computation. MPI consists of a standard set of API
calls that manage all aspects of communication and data transfer between
processors/nodes. MPI allows the coordination of a program running as
multiple processes in a distributed (not shared) memory environment,
yet is flexible enough to also be used in a shared memory system such as
a multi-core system.

Intel Support for MPI

Intel has both performance tuning software and its own MPI library. The
library is known as the Intel MPI Library, and is not specifically for any
brand of machine, or for that matter any particular interconnect. The
performance tuning tool that Intel developed to support optimized MPI
performance analysis is called the Intel Trace Analyzer and Collector.
Intel MPI Library
Intel created a version of MPI that eliminates a key drawback of MPI
librariesthe need to build a version of a program for each different
interconnect. The Intel MPI Library is possible because the library uses
298 Multi-Core Programming

the Direct Access Programming Library (DAPL) supplied by virtually

every interconnect vendor, plus Intel supplies its own library for generic
methods such as shared memory. This library allows a developer to
create an efficient program for all platforms in a single binary by linking
with one MPI that can automatically configure for the interconnect
present at run time. This changes MPI from its traditional source-level
compatibility only to also offer binary-level compatibility. This opens up
application developers to create executables of their programs that can
run on a dual-core processor desktop or a 4,096-node supercomputer,
while maintaining competitive performance with the old methods of
producing a separate build for each fabric.
The library offers a great deal of flexibility by allowing both users and
developers to select fabrics at run time. The library supports all MPI-1
features plus many MPI-2 features, including file I/O, generalized requests,
and preliminary thread support. The library is based on Argonne National
Laboratorys MPICH-2 release of the MPI-2 specification.
Intel offers this library for Linux. Microsoft recently started to offer a
version of an MPI library for Windows that also gives developers the ability to
have a single efficient binary as well. Figure 11.8 shows how an application is
linked with the Intel MPI library, which in turn accesses the DAPL layer.

MPI
Application

Intel MPI
c

Library

DALP DALP DALP DALP

TCP/IP
Shared Fallback
Memory (Sockets
Other Backup
c c
Inter- Interface)
InfiniBand Myrinet c
QsNet connection
Fabrics

Figure 11.8 Intel MPI Library Abstracts the DAPL-based Interconnects

 Chapter 11: Intel Software Development Products 299

Intel Trace Analyzer and Collector

The Intel Trace Analyzer and Collector allows a developer to analyze,
optimize, and deploy high-performance applications on clusters. The
collector interacts with an MPI application to collect information at run
time, and the analyzer is used to display the collected traces after the run to
allow analysis of the information. A developer can see concurrent behavior
of parallel applications through Timeline Views and Parallelism Displays, as
shown in Figure 11.9. The analyzer calculates statistics for specific time
intervals, processes, or functions. It also displays application activities,
event source code locations, and message passing along a time axis.

Figure 11.9 Timeline Views and Parallel Displays

Scalability is a key concern with any parallel program, and the analyzer
provides views that are particularly useful for a developer seeking to
enhance scalability. A user can navigate through trace data levels of
abstraction: cluster, node, process, thread, and function. The Detailed and
300 Multi-Core Programming

Aggregate Views, shown in Figure 11.10, allow examination of aspects of

application run-time behavior, grouped by functions or processes.

Figure 11.10 Detailed and Aggregate Views

For parallel application development on cluster systems, these offer

powerful capabilities and belong in any MPI developers toolkit. They
offer a great opportunity to understand MPI application behavior, which
in turns helps achieve high execution performance.

Key Points
Parallel programming is more natural than forcing our thinking into
sequential code streams. Yet, the change from this type of thinking that
developers have all been trained on means we all need to think
differently than we have for decades.
 Chapter 11: Intel Software Development Products 301

Assistance from the makers of developer tools can help us with

coding/expressing, debugging, tuning, and testing our parallel
applications.
Keep the following points in mind when using Intel software
development products:
Automatic parallelization by the compiler is cheap and easy to try,
and generally limited in how much it is likely to help.
Software-based Speculative Precomputation (SSP), also known as
Helper Threads, is a preview feature in Intels latest compilers.
SSP uses additional threads to hide memory latencies by
prefetching data into cache.
Intel Thread Checker can find errors in parallel programming
even when the errors are not causing the program to fail during
testing.
Gaining insight into the actual operation of an application may be
the most important way to help developers. The Intel VTune
Performance Analyzer with the Intel Thread Profiler are powerful
tools for gaining such insights.
The Intel VTune Performance Analyzer provides a non-intrusive
way of analyzing the performance of your application. The Intel
Thread Profiler is a view within the VTune tool that allows
programmers to profile multi-threaded programs.
The Intel Math Kernel Library and the Intel Integrated
Performance Primitives provide developers with high-
performance routines for math and multimedia routines,
respectively. Developers need to pay careful attention to the
interaction between threads in their local application and how
the libraries create and use threads.
Programming using message passing, such as with MPI, can lead
to highly scalable programs that can run on the largest computers
as well as new multi-core processors. Support for using MPI in an
application is provided by the Intel MPI Library. MPI-based
applications can be optimized using the Intel Trace Analyzer and
Collector.
The Intel Debugger, included as part of the Intel compiler
software distribution, provides multi-threading debugging
capabilities, including support for OpenMP.
302 Multi-Core Programming

These are complex and powerful tools. Describing all the features and
capabilities is beyond the scope of the book. For a more complete
discussion of all the different features and capabilities please refer to the
documentation included with the programs and stay up to date with the
latest information, which can be found at the Intel Software Network
Web site at www.intel.com/software.
 Glossary
64-bit mode The mode in which 64-bit applications run on platforms

with Intel Extended Memory 64 Technology (Intel EM64T). See
compatibility mode.
advanced programmable interrupt controller (APIC) The hardware
unit responsible for managing hardware interrupts on a computing
platform.
aliasing A situation in which two or more distinct references map to the
same address in memory or in cache.
alignment The need for data items to be located on specific boundaries
in memory. Misaligned data can cause the system to hang in certain
cases, but mostly it detrimentally affects performance. Padding helps
keep data items aligned within aggregate data types.
architecture state The physical resources required by a logical
processor to provide software with the ability to share a single set of
physical execution resources. The architecture state consists of the
general purpose CPU registers, the control registers, and the
advanced programmable interrupt controller (APIC). Each copy
of the architecture state appears to software as a separate physical
processor.
associativity The means by which a memory cache maps the main RAM
to the smaller cache. It defines the way cache entries are looked up
and found in a processor.

303
304 Multi-Core Programming

atomic Operations that are indivisible (hence, the reference to the

atom), meaning that all the operations must be completed or none of
them can be. Example atomic operations relevant to multi-core
programming include access to shared (global) memory locations,
database transactions, and any sequence of operations that, if
interrupted, would leave the software or the system in an unstable or
corrupted state.
barrier A synchronization mechanism that prevents forward progress by a
group of threads until all threads, or in some implementations, a certain
number of threads, reach the same point in their execution. Threads
will block until all threads in the group, or the threshold level has been
reached. Through this method, a thread from an operational set has to
wait for all or some number of other threads in that set to complete in
order to be able to proceed to the next execution step.
base address The starting address of a segment.
big-endian A way of storing values in memory that is favored by RISC
processors. In this scheme, for example, a two-byte integer whose
value is 0x0123 is stored in consecutive bytes in memory as 01 23.
The big-end (the most significant bits) are stored in the lower
addressed-byte, hence the name. See little endian.
cache coherency The need for caches on separate processors to contain
the same values with respect to a data item that is held in the cache
of more than one processor.
cache line The minimum amount of memory read by a processor from
RAM into its caches. The size of a cache line can be determined by
processor-specific calls. On recent IA-32 processors and those with
Intel EM64T, the cache line is 128 bytes.
canonical address A memory address that uses a special format and is
required for memory references. See effective address.
chip multiprocessing (CMP) A technology in which multiple processors
reside on a single die, are singly packaged, and utilize a single socket
of the platform.
Common Language Runtime (CLR) The virtual machine and execution
layer in Microsofts .NET environment.
coarse-grained multi-threading (Also referred to as switch-on-event
multi-threading) a type of threading in which the thread has full
control over processor resources for a specified threshold of timing
windows or number of processor cycles.
 Glossary 305

compatibility mode The operating mode in which 32-bit applications

can run on a 64-bit operating system on platforms with Intel EM64T.
concurrency The operational and execution methodology by which
common resources or CPUs simultaneously execute parallel tasks or
threads.
condition variable A mechanism that allows a thread to wait on some
condition or event to occur. Another thread will signal the waiting
thread once the condition occurs. Condition variables are almost
always used in conjunction with a mutex and a Boolean expression
that indicates whether or not the condition is met.
context switch The event that occurs when one thread stops executing
on a given CPU and yields control of the CPU to another thread.
During this event, the thread-specific state, including CPU registers,
instruction pointer, and stack pointer must be saved so that the
thread being evicted from the CPU may eventually be restored. This
event allows multiple processes to share a single CPU resource.
control registers Registers that are used to configure hardware to
operate in certain modes and perform specific actions. These
registers can be read and modified only by software at the highest
privilege levels, such as the operating system and device drivers.
convoying A common problem in lock-based synchronization where
multiple threads are blocked from executing while waiting to acquire
a shared lock. Convoys reduce concurrency and hence the overall
performance of multi-threaded programs
core The instruction execution portion of the processor, plus the caches
and interfaces to system buses.
cooperative multi-threading A form of multitasking where threads of
control voluntarily give up control of the CPU to other threads. In a
cooperative multi-threading scheme, threads are not pre-empted by
an external entity such as the operating system.
CPUID An assembly instruction that returns information on the runtime
processor. The information that it provides depends on the values of
parameters passed to it in various registers. The available data is
extended each time Intel modifies the processor architecture in an
important way.
critical section The part of a process where multiple threads overlap,
and that contains at least one shared resource that the various threads
may access. Only one thread is allowed to access the critical section
at any given time.
306 Multi-Core Programming

data decomposition The process of breaking down program tasks by

the data they work on rather than by the nature of the task.
data-level parallelism see data decomposition.
data race A condition where multiple threads attempt to access the
same memory location at the same time. In a data race, the value of
the memory location is undefined and generally incorrect. Data races
are generally avoided by synchronizing access to the shared memory
location.
deadlock A situation in which one thread (Thread A) is waiting for a
resource held by another thread (Thread B), while holding a resource
that is needed by Thread B. Since Thread A and Thread B are blocked
waiting for resources held by the other thread, the threads are locked
and no forward progress is made.
decomposition The process of breaking programs down into a series of
discrete tasks. There are three types of decomposition: functional,
data, and a variant of functional decomposition, called
producer/consumer.
Dekkers Algorithm A technique used in multi-threaded programming
for mutual exclusion that allows two threads to share a single
resource using shared memory.
double word On x86 architectures, a 32-bit data item.
dual-core is a term that describes a processor architecture in which two
processor cores are placed on the same physical die.
effective address A memory address that consists of two parts: a base
address and a displacement from that base address.
EM64T See Intel Extended Memory 64 Technology.
endian How numerical values are stored in bytes. See big-endian and
little-endian.
false sharing A problem that occurs when threads on two different
chips are both accessing the data item that are in the same cache
line. Each access requires both processors to update their caches. If
the updated item is used only by a single processor, all other
processors are still forced to update their caches despite the fact
they dont need to know about the change in data; hence the term,
false sharing.
fence A restraining mechanism, usually an instruction, that allows
synchronization among multiple attributes or actions in a system, and
ensures proper memory mapping from software to hardware memory
 Glossary 307

models. The fence instruction guarantees completeness of all pre-

fence memory operations and halts all post-fence memory operations
until the completion of fence instruction cycles.
fiber A thread that is scheduled and managed in user space. Also known
as green threads, or user-level threads.
fine-grained locking An operational locking mechanism where the
protection boundary is attributed to a single shared resource.
Multiprocessor and real-time kernels utilize fine-grained locking. This
increases concurrency in a system.
fine-grained multi-threading A type of threading in which the thread
switching takes place at an instruction cycle boundary.
flow dependence A type of data dependency in which one statement
depends on the value of a variable of the previous statement and
there is no redefinition of the variable between these two statements,
such as if a variable V is defined in statement S1 and later used at
statement S2 with no redefinition of V, then there exists a flow
dependency between S1 and S2.
functional decomposition A partitioning technique that subdivides a
program into different tasks based on the different independent
operations that it performs..
GDB (Gnu Debugger) A debugger that provides a number of
capabilities for debugging POSIX threads (Pthreads).
GDT See global descriptor table.
general-purpose register (GPR) A register that has no function pre-
assigned by the processor or the operating system. As such, GPRs can
be used for whatever purpose the software needs.
global descriptor table (GDT) A system table that can hold up to 8,192
entries that describe data items, such as segments, procedure entry
points, LDTs and the like. A given system must have a GDT.
GPR See general-purpose register.
hazard pointer A miniature garbage collector that handles pointers
involved in compare-exchange operations; called a hazard pointer,
because it presents a hazard to lockless algorithms.
horizontal multi-threading See simultaneous multi-threading (SMT).
Hyper-Threading Technology (HT Technology) Intels implementa-
tion of simultaneous multi-threading, in which multiple threads
execute simultaneously on the same processor core.
308 Multi-Core Programming

IA (Intel architecture) The family of Intel processors to which a given

chip belongs.
IA-32 (Intel architecture, 32 bits) The Intel architecture for processors

whose ILP is all 32-bits. It includes Intel Pentium 4 and Xeon
processors prior to the advent of Intel EM64T.
IA-32e (Intel architecture 32-bit extended) Shorthand for the Intel
architecture mode in which the instructions for Intel EM64T are in
use. It supports 32-bit software via compatibility mode and 64-bit
software via 64-bit mode.
IA-32 EL (Emulation Layer) A means of executing 32-bit IA-32 code on

an Intel Itanium processor.
IA-64 (Intel architecture, 64 bits) Shorthand for the 64-bit architecture
used in Intel Itanium processors.
inline C++ keyword that suggests to the compiler that it replace calls to
a function with the actual executable code for the function. This step
increases performance by eliminating the overhead of the function
call, but it can also detrimentally affect performance if overused by
increasing code size excessively.
ILP (integer-long-pointer) The programming model used by an operat-
ing system on a specific processor platform. It refers to the size in
bits of the integer, long integer, and pointer data types.
Intel EM64T See Intel Extended Memory 64 Technology.
Intel Extended Memory 64 Technology Provides 64-bit extensions to
the processor instruction set that enables IA-32e processors to
execute 64-bit operating systems and applications.
IPC (inter-process communication) The methodology by which
processes or computers exchange data with other processes or
computers using standard protocols.
Java Virtual Machine (JVM) A software interpreter that translates
precompiled Java bytecodes into machine instructions and executes
them on a hardware platform.
kernel thread A thread that is managed and scheduled by the kernel.
Kernel threads relieve the burden of scheduling and managing
threads from the programmer; however, kernel threads may require
additional overhead as they are managed by the operating system. As
a result, operations on them may require a system call. Most modern
operating systems have very efficient implementations that minimize
 Glossary 309

this overhead, making kernel threads the preferred threading

technique in most cases.
latency The preferred term for delay in the semiconductor industry.
LDT See local descriptor table
linear address An address that points to a byte in the systems linear
address space. On systems that dont use paging, the linear address is
the same as the physical address of the byte in RAM. On systems that
do use paging, this address must be converted to an actual physical
address by a series of lookups.
linear address space The memory that is addressable directly by the
processor. Under normal circumstances, this space would
correspond with 4 gigabytes of memory on 32-bit systems. This
address space is different from the physical address space, in that an
IA-32 system could be configured with less than 4 gigabytes of RAM.
In such a case, its linear address space remains 4 gigabytes, but its
physical address space is the lower number that corresponds to the
amount of RAM physically present.
little endian A way of storing values in memory that is favored by CISC
architectures, such as IA-32 processors. In this scheme, for example,
a two-byte integer whose value is 0x0123 is stored in consecutive
bytes in memory as 23 01 The little-end (the least important bits) are
stored in the lower addressed-byte, hence the name. See big endian.
live lock The hazard status that occurs when threads continually conflict
with each other and back off.
load balancing The distribution of work across multiple threads so that
they all perform roughly the same amount of work.
local descriptor table A system table that can hold up to 8,192 entries.
These entries describe system data items, such as segments,
procedure entry points, and the like. A system can have zero or more
Local Descriptor Tables.
lock A mechanism for enforcing limits on access to a shared resource in
an environment that has many threads of execution. Locks are one
way of enforcing concurrency control policies.
logical processor The hardware interface exposed by processors with
Hyper-Threading Technology that makes it appear, from softwares
perspective, that multiple processors are available. This is accom-
plished by duplicating the architecture state of the processor,
including the CPU register set and interrupt control logic. Logical
processors share a single set of physical execution resources.
310 Multi-Core Programming

loop scheduling In OpenMP, a method of partitioning work done in a

loop between multiple threads. There are four loop scheduling types,
including static, dynamic, guided, and runtime.
memory cage A relaxed memory model idiom that starts with an
acquire fence and ends in a release fence. Memory cages keep things
that are inside from getting out, and not vice versa.
memory latency The delay caused by accessing RAM memory. This
delay arises in large part because RAM chips run at one tenth the
clock speed of most modern processors.
message A special method of communication to transfer information or
a signal from one domain to another.
micro-ops The smallest executable unit of code that a processor can
run. IA-32 instructions are translated into micro-ops before they are
executed.
MMX Technology An extended instruction set introduced by Intel to
improve performance in multimedia applications.
model-specific register (MSR) Special registers that vary from one
processor generation to the next. They contain data items that are
used by the operating system and the processor for memory
management, performance monitoring, and other system functions.
monitor A simplistic, abstracted synchronization methodology that
guarantees mutual exclusion of internal data and has thread
synchronization capabilities. A critical section gets included in a
monitor to allow a thread exclusive access to internal data without
interference from other threads.
MSR See model-specific register.
multi-core is a term that describes a processor architecture in which two
or more processor cores are placed on the same physical die.
multiple-processor system-on-a-chip See chip multiprocessing
(CMP).
multitasking A technique used by operating systems that allow users to
run multiple processes, or applications, simultaneously.
multi-threading A technique used to run multiple threads of execution
in a single process or application.
mutex The mechanism by which threads acquire control of shared
resources. A mutex is also referred to as mutual-exclusion
semaphore. A mutex has two states, locked and unlocked.
 Glossary 311

mutual exclusion A technique for synchronization among threads that

limits execution in a particular critical section of code to a single
thread of execution.
mutual exclusion semaphore See mutex.
.NET Microsofts managed execution environment.
non-blocking algorithm An algorithm designed not to use locks. The
defining characteristic of a non-blocking algorithm is that stopping a
thread does not prevent the rest of the system from making progress.
Non-Uniform Memory Access (NUMA) An architecture that physically
links two or more SMPs, where one SMP can access memory of
another SMP. As the name NUMA suggests, not all processors have
equal access time to the memory. When cache coherency is
preserved for NUMA architecture, it is called cc-NUMA.
NUMA See Non-Uniform Memory Access.
on-chip multiprocessing See chip multiprocessing (CMP).
OpenMP An application programming interface that provides a platform-
independent set of compiler pragmas, directives, function calls, and
environment variables that explicitly instruct the compiler how and
where to use parallelism in an application.
padding Unused bytes placed in a structure or other aggregate data
object to assure that all fields are properly aligned.
PAE See physical address extensions.
page directory base register A register that contains the base address
of the system page directory. This register is generally collocated in
control register CR3.
page directory entry An entry in the systems page directory table that
contains detailed data about the status of a given memory page, such
as whether its present in memory, and about what access rights can
be granted to it.
page directory pointer table In 64-bit mode, this second look-up table is
consulted during resolution of linear addresses to physical addresses. It
occurs after the PML4 and before the page directory table.
page map level 4 table The first look-up table used in resolving linear
addresses in 64-bit mode to physical addresses.
page size extensions Technology that enables page sizes beyond the
default 4 kilobytes. Using page size extensions, pages can be
2 megabytes or 4 megabytes.
312 Multi-Core Programming

page size extensions (36-bits) An alternative to PAE for extending

addresses to 36-bits on IA-32 architectures.
parallelism The operational and execution methodology by which
different resources or CPUs execute the same task simultaneously.
PDBR See page directory base register.
physical address The actual location of an item in the physical address
space.
physical address extensions A method of extending 32-bit addresses
to 36 bits on IA-32 architectures. On platforms with Intel EM64T, PAE
enables similar extensions that are implemented differently.
physical address space The range of addresses that the processor can
generate on the memory bus. Hence, it reflects the total amount of
addressable physical memory on a given system.
PML4 See page map level 4 table.
POSIX thread (Pthread) A portable, standard threading API that is
supported on a number of different operating systems, including
many different flavors of Unix, MacOS, and Microsoft Windows.
priority ceilings A method for raising the priority of a thread when it
acquires a lock. A priority ceiling is the maximum priority level of any
thread that needs to access a critical section of code. The thread that
acquires a lock that has a priority ceiling value immediately runs at
that priority level. This technique is used to avoid the problem of
priority inversion.
priority inheritance A method of allowing a lower priority thread to
inherit the priority of a high priority thread when the low priority
thread holds a lock that is needed by the high priority thread. This is
commonly used to avoid the problem of priority inversion.
priority inversion A threading problem observed in priority based
schedulers where a low priority thread holds a lock that is required
by a high priority thread. Meanwhile, a medium priority thread is
running, preventing the low priority thread from releasing the lock,
thus starving the high priority thread. This bug was encountered on
the Mars Pathfinder mission.
privilege level The level of permissible activities as enforced by the
processor. Privilege level 0 is the highest level of privilegeall
processor instructions can be executed at this privilege level. Most
applications running on IA-32 architectures run at privilege level 3.
 Glossary 313

Certain instructions that change the way the processor functions

cannot be executed from this privilege level.
preemptive multi-threading A thread-based scheduling technique
where the currently running thread is stopped by an external entity,
usually the operating system. The current thread loses control of the
CPU, and another thread is allowed to execute. The process of
switching between threads is known as a context switching.
prefetching Loading data items or instructions into cache prior to the
processors need for them. Prefetching prevents processor stalls by
making sure that needed data is always in cache. Modern Intel IA-32
processors support prefetching in hardware and software.
process A process is a program in execution. It contains a main thread of
execution, as well as an address space, and other resources allocated
to it by the operating system. Processes are sometimes referred to as
heavy-weight processes.
processor affinity The preference for a thread to run on a given
processor.
processor order A semi-relaxed memory consistency model. This
method maintains the correctness of consistency of memory read-
write sequences.
producer/consumer (P/C) decomposition A common form of
functional decomposition where the output of one task, the
producer, becomes the input to another, the consumer..
PSE See page size extensions
PSE-36 See page size extensions (36-bits)
Pthread See POSIX thread
read-write lock A lock that allows simultaneous read access to multiple
threads but limits the write access to only one thread.
recursive lock A lock that is called recursively by the thread that
currently owns the lock.
register Registers are special, high-speed locations in the processor
where data items and addresses are placed for use by the processor in
executing a specific instruction.
register file The collection of all the registers available on a specific
processor.
relaxed consistency A memory consistency model that maintains
memory to be consistent only at certain synchronization events and
314 Multi-Core Programming

ensures maintenance of memory write operations by following

consistency constraints. Relaxed consistency helps reduce the cost of
memory access by hiding the latency of write operations.
quadword On IA-32 architectures, a 64-bit data item.
register pressure A situation in which software requires more registers
than are presently available, which leads to excess swapping of data
in and out of registers, thereby reducing performance.
Savage benchmark A Fortran benchmark designed by Bill Savage to
exercise a systems floating-point arithmetical capabilities.
segment A block of memory used for a discrete task by a program or the
operating system.
segment descriptor An entry in a descriptor table that contains
important data regarding a specific segment.
segment selector The part of a logical address that serves as a reference
to a segment descriptor in a descriptor table.
serializing event An instruction or action that causes the processor to
cease all speculative and out-of-order execution and discard the
results of any instruction executed but not retired. The processor
then resumes processing at the current instruction. These events
occur because certain execution aspects, such as precision of
floating-point calculations, have changed, generally at the request of
the running program.
segment override The act of loading a specific value in a segment base
register, rather than employing the default value. In 64-bit mode,
segment overrides can be performed on the FS and GS segment
registers only.
semaphore A special type of variable used for synchronization.
Semaphores can only be accessed by two operations, wait and signal.
A semaphore is an extension of a mutex, and allows more than one
thread in a critical section.
sequential consistency When, at any given instant in time in a
sequential program, memory has a well defined state.
single-issue, single-thread A baseline processor threading model that
does not exploit any parallelism.
SIMD The acronym stands for single instruction, multiple data items, a
technology in which a single arithmetic operation is performed on
multiple data items at one time. It is frequently useful for the
arithmetic performed in multimedia and imaging applications. Intels
 Glossary 315

Streaming SIMD Extensions (SSE) family of technologies uses SIMD

extensively.
simultaneous multi-threading (SMT) A processor multi-threading
model that allows threads to compete for shared available resources
and enhance processor utilization.
soft affinity The policy used by Microsoft Windows to select a
processor for execution of a thread. This policy suggests to the
Windows scheduler that threads should, as much as possible, run on
the processor on which they ran previously.
spin wait A tight or time-delayed loop-based locking mechanism used
for synchronization. Spin waits allow a thread to wait for something
else to happen instead of calling an interrupt.
SSE (Streaming SIMD Extensions) Extensions to IA-32 processors designed
for fast performance of routine tasks, especially in multimedia and
imaging applications. SSE makes extensive use of SIMD.
SSE2 (Streaming SIMD Extensions 2) The second generation of SSE
instructions introduced by Intel with the Pentium 4 processor in
2001. It added support for 64-bit double-precision arithmetic and
included cache management instructions, among other features.
SSE3 (Streaming SIMD Extensions 3) The third generation of SSE
instructions introduced with the Prescott generation of Pentium 4
processors in 2004. Among other features, these 13 instructions add
capabilities for performing arithmetic within a single XMM register
and efficiently converting floating-point numbers to integers.
synchronization The process by which two or more threads coordinate
their activities and allow threads to coexist efficiently. There are four
distinct means available for synchronizing threads: mutexes,
condition variables, read/write locks, and semaphores.
taskqueuing An extension to OpenMP by Intel that allows programmers
to parallelize special control functions such as recursive functions,
dynamic tree searches, and pointer chasing while loops.
task state segment All the data the processor needs to manage a task is
stored in a special of segment, known as the task state segment.
thread The minimal schedulable execution entity. Threads contain an
instruction pointer to the instruction stream that the thread executes,
CPU state information, and a stack pointer. Depending on the platform,
additional information may be kept by the operating system or hardware
platform. Threads are sometimes called lightweight processes.
316 Multi-Core Programming

thread pool A collection of worker threads that are used to perform

independent units of work on demand. Thread pools reduce the
overhead of thread creation/destruction.
thread safe A property of a section of code that determines whether or
not multiple threads may simultaneously execute that block of code.
A function is considered to be thread safe if and only if the function
returns the proper results when multiple threads are calling the
function at the same time.
thunking A form of backwards-compatibility mechanism used by
Microsoft in Windows to enable older binaries to run on modern
versions of Windows.
translation look-aside buffer (TLB) An on-chip cache that holds page-
lookup information.
TSS See task state segment.
uniform memory access (UMA) UMA is also referred to as symmetric
multiprocessor (SMP) or cache coherent UMA (CC-UMA) architec-
ture. With UMA, all the processors are identical and have equal
access time to memory. Cache Coherent means, if one processor
updates a location in shared memory, all the other processors know
about the update.
virtual 8086 A method of executing 16-bit programs on IA-32
processors. No longer used today, but supported on IA-32 processors
for backward compatibility. It is not supported in IA-32e
Compatibility mode.
virtual machine monitor (VMM) A virtualization layer between a host
system and virtual machines, where the virtual machine is the
container of operating systems. VMM is also referred to as a
hypervisor.
word The basic, default amount of data that a given architecture uses. It
is generally as wide as an integer and the size of the default address.
 References
Books and Articles
Abrams, Brad. 2004. .NET Framework Standard Library Annotated
Reference, Volume 1: Base Class Library and Extended Numerics
Library. Redmond, WA: Microsoft Press.
Alagarsamy, K. 2003. Some Myths About Famous Mutual Exclusion
Algorithms, ACM SIGACT News, Vol. 34, No. 3, September, 2003.
Andrews, Gregory R. 2000. Foundations of Multithreaded, Parallel, and
Distributed Programming. Boston, MA: Addison-Wesley.
Ang, Boon, Derek Chiou, Larry Rudolph, and Arvind. 1996. Message
Passing Support for Multi-grained, Multi-threading, and Multi-tasking
Environments. MIT Laboratory for Computer Science, Computation
Structures Group, Memo 394.
Barham, Paul, Boris Dragovic, Keir Fraser, Steven Hand, Tim Harris, Alex
Ho, Rolf Neugebauery, Ian Pratt, and Andrew Warfield. 2003. Xen and
the Art of Virtualization. University of Cambridge Computer Laboratory,
15 JJ Thomson Avenue, Cambridge, UK, CB3 0FD, SOSP03, October
1922, 2003, Bolton Landing, New York, USA.
Barney, Blaise. Introduction to Parallel Computing. Lawrence Livermore
National Laboratory, Livermore Computing. Available at:
http://www.llnl.gov/computing/tutorials/parallel_comp/.

317
318 Multi-Core Programming

Barney, Blaise. 2006. POSIX Threads Programming. Lawrence Livermore

National Laboratory. Available at: http://www.llnl.gov/computing/
tutorials/pthreads/
Blumofe, Robert D., Christopher F. Joerg, Bradley C. Kuszmaul, Charles
E. Leiserson, Keith H. Randall, and Yuli Zhou. 1995. Cilk: An Efficient
th
Multithreaded Runtime System. Proceedings of the 5 ACM SIGPLAN
Symposium on Principles and Practice of Parallel Programming
(July):207216.
Brinch Hansen, Per. 1972. Structured Multiprogramming. Communications
of the ACM, 15(7):574578.
Bulpin, James Roy. 2004. Operating System Support for Simultaneous
Multithreaded Processors. PhD thesis, Kings College, University of
Cambridge, September.

Butenhof, David R. 1997. Programming with POSIX Threads. Boston,
MA: Addison-Wesley Professional.
Chandra, Rohit, Leonardo Dagum, Dave Kohr, Dror Maydan, Jeff
McDonald, and Ramesh Menon. 2001. Parallel Programming in
OpenMP. San Francisco, CA: Morgan Kaufmann Publishers.
Culler, David E., Jaswinder Pal Singh. 1999. Parallel Computer
Architecture A Hardware/Software Approach. San Francisco, CA:
Morgan Kaufmann.
Dijkstra, Edsger W. 1968. The Structure of the THE Multiprogramming
System. Communications of the ACM, 11(5):341346.
Frigo, Matteo and Steven G. Johnson. 1997. Fastest Fourier Transform in
the West. Massachusetts Institute of Technology. Technical Report
MIT-LCS-TR-728 (September).
Garcia-Rosello, Emilio, Jose Ayude, J. Baltasar Garcia Perez-Schofield, and
Manuel Perez-Cota. 2002. Design Principles for Highly Reusable
Concurrent Object-Oriented Systems. Computer Sciences Department,
University of Vigo, Spain. Vol. 1, No. 1, May-June.
Gerber, Richard, Aart J. C. Bik, Kevin B. Smith, and Xinmin Tian. 2006.
The Software Optimization Cookbook, Second Edition. Hillsboro,
OR: Intel Press.
Goodman, James R., Mary K. Vernon, and Philip J. Woest. 1989. Efficient
Synchronization Primitives for Large-Scale Cache-Coherent
Multiprocessors. ACM, 1989 0-89791-300-0/89/0004/0064
 References 319

Grama, Ananth, Anshul Gupta, George Karypis, and Vipin Kumar. 2003.
Introduction to Parallel Computing. Boston, MA: Addison-Wesley.
Hennessy, John L. and David A. Patterson. 2003 Computer Architecture
A Quantitative Approach. San Francisco, CA: Morgan Kaufmann.
Hill, Mark D. 1998. Multiprocessors Should Support Simple Memory
Consistency Models. IEEE Computer (August), 31(8):2834.
Hoare, C.A.R. 1974. Monitors: An Operating System Structuring Concept.
Communications of the ACM, 17(10):549557.
Holt, Bill. 2005. Moores Law, 40 years and Counting Future Directions
of Silicon and Packaging. InterPACK 05, Heat Transfer Conference.
Holub, Allen. 2000. Taming Java Threads. Berkeley, CA: Apress.
Hughes, Cameron and Tracey Hughes. 2004. Dividing C++ Programs
into Multiple Threads. Boston, MA: Addison Wesley.

Intel Corporation. 2003. Intel Hyper-Threading Technology, Technical
Users Guide. Santa Clara, CA: Intel Corporation.
______. 2005. IA-32 Intel Architecture Optimization Reference Manual.
Available at: http://www.intel.com/
______. 2006a. Intel Itanium Architecture Software Developers
Manual, Volume 1: Application Architecture, Volume 2: System
Architecture, Volume 3: Instruction Set Reference. Available at:
http://www.intel.com/
______. 2006b. IA-32 Intel Architecture Software Developers Manual,
Volume 1: Basic Architecture, Volume 2A-2B: Instruction Set
Reference, Volume 3: System Programming Guide. Available at:
http://www.intel.com/
______. 2006c. Intel Processor Identification and the CPUID Instruction,
Application Note 485. Available at: http://www.intel.com/
Kleiman, Steve, Devang Shah, and Bart Smaalders. 1996. Programming
with Threads. Upper Saddle River, NJ: Prentice Hall.
Kongetira, Poonacha, Kathirgamar Aingaran, and Kunle Olukotun. 2005.
Niagara: A 32-way Multithreaded SPARC Processor. IEEE Micro
(March/April), 25(2):2129.
Kubiatowicz, John David. 1998. Integrated Shared-Memory and Message-
Passing Communication in the Alewife Multiprocessor. PhD Thesis,
Massachusetts Institute of Technology, February.
320 Multi-Core Programming

Lea, Doug. 1997. Concurrent Programming in Java Design Principles

and Patterns. Boston, MA: Addison-Wesley.
Lo, Jack L., Susan J. Eggers, Joel S. Emer, Henry M. Levy, Rebecca L.
Stamm, and Dean M. Tullsen. 1997. Converting Thread-Level
Parallelism to Instruction-Level Parallelism via Simultaneous
Multithreading. ACM Transactions on Computer Systems, Vol. 15,
No. 3, August,
Mattson, Tim. Nuts and Bolts of Multithreaded Programming. Santa Clara,
CA: Intel Corporation. Available at: http://www.intel.com.
Mattson, Tim, Beverly Sanders, and Berna Massingill. 2004. Patterns for
Parallel Programming. Boston, MA: Addison-Wesley Professional.
Michael, Maged. 2004. Hazard Pointers: Safe Memory Reclamation for
Lock-Free Objects. IEEE Transactions on Parallel and Distributed
Systems (June), 15(6):491504.
Nutt, Gary J. 2000. Operating Systems A Modern Perspective. Boston,
MA: Addison-Wesley.
Redstone, Joshua A., Susan J. Eggers, and Henry M. Levy. 2000. An
Analysis of Operating System Behavior on a Simultaneous
Multithreaded Architecture. Proceedings of the 9th International
Conference on Architectural Support for Programming Languages
and Operating Systems, November.
Reek, Kenneth A. 2002. The Well-Tempered Semaphore: Theme with
Variations. ACM SIGCSE Bulletin (March), 34(1):356359.
Reeves, Glenn E. 1998. Re: What Really Happened on Mars? Risks-
Forum Digest (January), 19(58).
Seng, John S., Dean M. Tullsen, and George Z. N. Cai. 2000. Power-
Sensitive Multithreaded Architecture. IEEE, 0-7695-0801-4/00
Shen, John Paul and Mikko H. Lipasti. 2005. Modern Microprocessor
Design Fundamentals of Superscalar Processors. New York:
McGraw-Hill.
Smith, James E. and Ravi Nair. 2005. Virtual Machines Versatile
Platforms for Systems and Processes. San Francisco, CA: Morgan
Kaufmann.
Stokes, Jon. 2002. Introduction to Multithreading, Superthreading and
Hyperthreading. Ars Technica, October.
 References 321

Sutter, Herb and James Larus. 2005. Software and the Concurrency
Revolution. Microprocessor, Vol. 3, No. 7, September.
Ungerer, Theo, Borut Robi, and Jurij ilc. 2003. A Survey of Processors
with Explicit Multithreading. ACM Computing Surveys, Vol. 35, No. 1,
March.
Vahalia, Uresh. 1995. UNIX Internals: The New Frontiers. Upper Saddle
River, NJ: Prentice Hall.
Wadleigh, Kevin R. and Isom L. Crawford. 2000. Software Optimization
for High Performance Computing. Indianapolis, IN: Prentice Hall
PTR.
Wisniewski, Robert W., Leonidas I. Kortothanassis, and Michael L. Scott.
1995. High Performance Synchronization Algorithms for
Multiprogrammed Multiprocessors. Department of Computer
Science, University of Rochester, Rochester, NY.

Tools and Web Sites

Cluster computing resources: http://www.buyya.com/cluster/
Intel Software Web site: http://www.intel.com/software/
Message Passing Interface (MPI) Forum: http://www.mpi-forum.org
Message Passing Interface (MPI) training:
http://www.mhpcc.edu/training/workshop/mpi/
POSIX Threads Programming. Lawrence Livermore National Laboratory:
http://www.llnl.gov/computing/hpc/training/
pthreads-win32. An open source version of Pthreads for Windows. The
library is actively maintained and is under the stewardship of Red
Hat, Inc: http://sourceware.org/pthreads-win32/
 About the Authors
Shameem Akhter is a platform architect at Intel, focusing on single socket multi-
core architecture and performance analysis. He has also worked as a senior software
engineer with the Intel Software and Solutions Group, designing application
optimizations for desktop and server platforms. Shameem holds a patent on a
threading interface for constraint programming, developed as a part of his masters
thesis in computer science.
Jason Roberts is a senior software engineer at Intel Corporation. Over the past 10
years, Jason has worked on a number of different multi-threaded software products
that span a wide range of applications targeting desktop, handheld, and embedded
DSP platforms.

Major Contributors
James Reinders is a senior engineer who joined Intel in 1989 and has contributed
to projects including the worlds first TeraFLOP supercomputer, as well as
compilers and architecture work for Intel processors. He is currently director of
business development and marketing for Intels Software Development Products
group and serves as the chief evangelist and spokesperson. James is also the author
of the book VTune Performance Analyzer Essentials.
Arch D. Robison has been a Principle Engineer at Intel since 2000. Arch received
his Ph.D. in computer science from the University of Illinois. Prior to his work at
Intel, Arch worked at Shell on massively parallel programs for seismic imaging. He
was lead developer for the KAI C++ compiler and holds five patents on compiler
optimization.
Xinmin Tian holds a Ph.D. in computer science and leads an Intel development
group working on exploiting thread-level parallelism in high-performance Intel C++
and Fortran compilers for Intel Itanium, IA-32, Intel EM64T, and multi-core
nd
architectures. Xinmin is a co-author of The Software Optimization Cookbook, 2
Edition.
Sergey Zheltov is a project manager and senior software engineer at the Advanced
Computer Center of Intels Software Solutions Group. He holds an M.S. degree in
theoretical and mathematical physics. Sergeys research includes parallel software
and platform architecture, operating systems, media compression and processing,
signal processing, and high-order spectra.
 Index
threading for .NET Framework, 107, 109,
A 110, 111, 112, 113, 114, 115, 116, 117,
118, 119, 120
ABA problem, 188, 189, 190 threading for POSIX, 120, 121, 122, 123,
Abort( ) method, 110, 111, 112 124, 125, 126, 127, 128, 129, 130,
ACPI (Advanced Configuration and Power 131, 132
Interface), 270 threading for Windows, 75, 76, 77, 78, 79,
acquire( ) operation, 63, 64 80, 81, 82, 83, 84, 85, 86, 87, 88, 90, 91,
92, 93, 94, 95, 96, 97, 98, 99, 100, 101,
AddEntryToTraceBuffer( ) method, 222
102, 103, 104, 105, 106
$ADDRESS tracepoint keyword, 226
threads and, 23, 24
affinity. See processor affinity
application programming models
AfxBeginThread( ) function, 79 concurrency in software, 2, 3, 4, 5, 19
AfxEndThread( ) function, 79 debugging in Windows, 224, 225, 227, 228,
algorithms, non-blocking, 186, 187, 229, 230, 231, 232
188, 189, 190, 191 designing with debugging in mind, 216,
ALU (Arithmetic Logic Unit), 239 217, 218, 219
Amdahl, Gene, 14 Intel software development and, 276, 277
Amdahls law legacy applications, 76
Gustafson's law and, 18, 19 parallel programming problems, 210, 211
overview, 14, 15, 16, 17, 18 threads and, 32, 33
work distribution and, 294 arbiters, 263, 264
anti-dependence, 138, 139 asynchronous I/O, 94, 177
APIC (Advanced Programmable Interrupt Atomic directive (OpenMP), 282
Controller), 242, 266 atomic pragma (OpenMP), 157, 159, 160
APIs auto-parallelization feature (Intel), 283, 284,
MPI programming and, 297 285, 286, 287

323
324 Multi-Core Programming

B CGMT (coarse-grained multi-threading),

248, 249
bandwidth chipsets, 237
cache usage and, 287 Cilk scheduler, 174
memory issues, 194, 198
CLR (Common Language Runtime)
barrier mechanism, 72, 73 terminating threads, 110
barrier pragma (OpenMP), 152, 153, 155 threading APIs, 75, 107
Barsis, E., 18 threads and, 24, 33
_beginthreadex( ) function, 78, 83 CMP (chip multiprocessing), 9, 250
binary semaphores, 61 CMT (chip multi-threading), 30, 250
blocking conditions, 95 code reviews, 218, 219
boundaries, threads and, 68 commands, groups of threads
branch prediction, 255 and, 235
break command (GDB), 233, 234 communication
break statement, 137 defined, 42
breakpoints inter-processor, 266, 268
GDB and, 233, 234, 235 managing threads and, 81, 82, 83, 84, 85,
Windows filters, 226, 227 86, 87
broadcast( ) operation, 66 messages and, 68, 69, 70
signaling POSIX threads, 125, 126, 127, 128,
bt (backtrace) command (GDB), 235
129, 130, 131, 132
comparison operation (OpenMP),
C
136, 137
C++ language compilation
.NET Framework and, 107 in Windows, 105, 106
memory reclamation, 190 just-in-time, 33
multi-threaded libraries, 106 memory issues, 194
OpenMP and, 164 OpenMP standard, 162, 164, 165
parallel programming problems, 211 Pthreads, 132
reduction operators, 148, 149 thread-safe libraries, 193
cache-oblivious blocking technique, 195 compiler switches
caching memory OpenMP, 164
multi-threading and, 12, 13 Windows, 105
optimizing, 287 compute threads, 173
parallel programming problems, 172, 190, concurrency in software, 2, 3, 4, 5
200, 201, 202, 203, 204, 205, 206, 207, condition synchronization, 54
208, 209, 210, 211 condition variables
Call Graph feature (VTune analyzer), 294 Pthreads and, 125, 126, 127, 128, 129
$CALLER tracepoint keyword, 226 synchronization primitives, 66, 67, 68, 70
$CALLSTACK tracepoint keyword, 226 ConvertThreadToFiber( ) function, 101, 102
captureprivate clause (OpenMP), 160 convoying behavior, 173, 181
CAS (Compare and Swap) operation, cooperative multi-threading, 29
64, 187, 191 copyin clause (OpenMP), 155
 Index 325

copyprivate clause (OpenMP), 155 deadlocks

core debugging, 278
Itanium architecture, 262 Intel Thread Checker, 278
overview, 257, 258, 259 nested critical sections and, 158
CPU (Central Processing Unit), 239 parallel programming constructs,
CPU clock frequency, 7 57, 59
parallel programming problems, 177, 178,
CreateEvent( ) method, 83, 84
179, 180
CreateFiber( ) function, 101, 102
thread messaging and, 70
CreateFile( ) function, 106
debugging
CreateMutex( ) function, 90 applications in Windows, 224, 225, 226,
CreateSemaphore( ) function, 88 227, 228, 229, 230, 231, 232
CreateThread( ) function, 76, 77, 78, 79, GDB and, 232, 233, 234, 235
83, 106 general techniques, 215, 216, 217, 218, 219,
critical pragma (OpenMP), 157 220, 221, 222, 223, 224
critical sections Intel Debugger, 288, 289
defined, 88 Intel software development products
named, 157, 158 and, 277
nested, 157, 158 OpenMP and, 165, 166, 167
parallel programming constructs and, thread-safe libraries and, 193
56, 57 decomposition
performance and, 157 data, 39
Win32 API and, 87 data flow, 40, 41
CRITICAL_SECTION data structure, 91 defined, 38
CriticalSection mechanism, 73, 74 geometric decomposition pattern, 44
implications of, 41, 42
D task, 38, 39, 41, 42
Decrement( ) method, 119
DAPL (Direct Access Programming DEE (dynamic execution engine), 256
Library), 298
default clause (OpenMP), 141, 166
data decomposition, 39, 41
Dekkers Algorithm, 205
data dependence, threading loops and, 138,
DeleteCriticalSection( ) function, 91
139, 140
DeleteFiber( ) function, 102
data flow decomposition, 40, 41
_declspec( ) directive, 202
data organization, performance and,
212, 213 dies, 239, 269
data streams, 5 Dijkstra, Edsger, 60, 61
data-level parallelism, 39 divide and conquer pattern, 43
data-race conditions DLLs (dynamically linked libraries), 278
debugging, 222, 278 DMI (direct media interface), 239
Intel Thread Checker, 278 domains, messages and, 68
OpenMP and, 140, 141, 166 double-check idiom, 210
parallel programming problems, 174, DTLB (Data Translation Lookaside
176, 177 Buffer), 257
326 Multi-Core Programming

dynamic execution. See ILP filters, breakpoint, 226, 227

dynamic scheduling scheme (OpenMP), 143, fine-grained locking, 184, 185
145, 146 firstprivate clause (OpenMP), 141, 142, 155,
156, 167
E flow dependence, 138
Floyd-Steinberg error weights, 47
ECC (error-correction code), 263
Flynn's taxonomy, 5, 6
Eckert, J. Presper, 1
frequency, 269, 270
EDVAC (Electronic Discrete Variable
Automatic Computer), 1 FSB (front side bus)
_endthreadex( ) function, 78 multi-core processors, 257, 262
single-core processors, 238, 239,
Enhanced Intel SpeedStep
241, 242
Technology, 270
$FUNCTION tracepoint keyword, 226
Enter( ) method, 118
Function( ) routine, 93, 94
EnterCriticalSection( ) function, 91
functional units, 239, 242, 257
environment variables, 163, 164
EPIC (Explicitly Parallel Instruction
Computing) G
multi-core processing and,
252, 262 garbage collection, 189, 190, 191
superscalar and, 244, 245, 246 GDB (GNU Project Debugger), 222, 232, 233,
error diffusion algorithm, 45, 46, 47, 48, 49, 234, 235
50, 51 geometric decomposition pattern, 44
EVENT_MODIFY_STATE field, 86 GetCurrentFiber( ) function, 102
Exchange( ) method, 119, 120 GetDevicePowerState API, 268
execution core, 257 GetExitCodeThread( ) function, 93
Executor class (Java), 174 GetFiberData( ) function, 101
Exit( ) method, 118 GetMaxThreads( ) method, 113
ExitThread( ) function, 77, 78 GetMinThreads( ) method, 113
GetProcessAffinityMask( ) function, 98
F GetProcessPriorityBoost( ) function, 95
false sharing, 13, 200, 201, 202, 203 GetSystemPowerStatus API, 268
FCFS (First-Come-First-Serve) model, 61 GetThreadPriority( ) function, 96
fence mechanism, 71, 206, 207, 210 GetThreadPriorityBoost( ) function, 95
fetch-and-op operation, 187, 188, 191 GMCH (Graphics Memory Controller
Hub), 239
FGMT (fine-grained multi-threading),
248, 249 goto statements, 61, 137
fibers guided scheduling scheme (OpenMP), 143,
145, 146
defined, 27
user-level threading and, 100, 101, 102, GUIs (Graphical User Interfaces), 2
103, 104 Gustafson's law, 18, 19
FIFO (First-In, First-Out), 66
 Index 327

H instruction streams, 5
Integrated Performance Primitives (Intel), 290
HAL (hardware abstraction layer), 26
Intel Core Duo processor, 6, 252
hardware
Intel Debugger, 288, 289
message processing, 70
Intel MPI Library, 298
threads and, 22, 29, 30, 248, 249, 250,
251, 252 Intel NetBurst microarchitecture, 251, 255
hash tables, 184, 185 Intel software development products
compilers, 282, 283, 284, 285, 286, 287
hazard pointers, 190
Intel Debugger, 288, 289
hint@pause instruction, 265
Intel Thread Building Blocks, 292
HLO (High Level Optimizations), 284
Intel Thread Checker, 141, 176, 277, 278,
HT (Hyper-Threading) Technology 279, 280, 281
Amdahl's law, 17, 18 Intel Thread Profiler, 276, 277, 295, 296
concurrent processing and, 29 Intel VTune Performance Analyzer, 141,
differentiating multi-core architectures, 169, 276, 277, 292, 293, 294, 295
10, 11 libraries, 289, 290, 291
false sharing, 202 MPI programming, 296, 297, 298,299, 300
overview, 8 overview, 276, 277
runnable threads, 173 Intel Thread Building Blocks, 292
thread priorities, 96
Intel Thread Checker
threading APIs, 99, 100
data-race conditions, 141, 176
threading on, 252, 253, 254, 255, 256, 257
overview, 277, 278, 279, 280, 281
Intel Thread Profiler, 276, 277, 295, 296
I
Intel Trace Analyzer and Collector, 299, 300
I/O APIC, 242 Intel VTune Performance Analyzer
I/O threads, 173 data-race conditions, 141
IA-32 architecture, 205, 206, 207, 211, 212 investigating applications, 276
ICH (I/O Controller Hub), 239 overview, 292, 293, 294, 295, 296
IHA (Intel Hub Architecture), 239 performance and, 169
tuning and, 277
ILP (instruction-level parallelism)
defined, 248 interface unit, 242
goal of, 7 interleaving, 154, 155
Itanium architecture and, 262 Interlocked API, 91, 92
multi-core processors and, 249 Interlocked class (.NET), 119, 120
implementation-dependent threading, 73, 74 InterlockedCompareExchange( ) function, 92,
Increment( ) method, 119 187, 188
INFINITE constant, 84 InterlockedCompareExchange64( )
function, 92
info command (GDB), 233
InterlockedCompareExchangePointer( )
initial threads, 32
function, 92
InitializeSListHead( ) function, 92
InterlockedDecrement( ) function, 92
InitializeTraceBuffer( ) method, 222
InterlockedExchange( ) function, 92
instruction pointers (IP), 255
328 Multi-Core Programming

InterlockedExchange64( ) function, 92 L
InterlockedExchangeAdd( ) function, 92
lastprivate clause (OpenMP), 141, 142,
InterlockedExchangePointer( ) function, 92
155, 167
InterlockedFlushSList( ) function, 92
latency hiding, 10, 11
InterlockedIncrement( ) function, 92
LeaveCriticalSection( ) function, 91
InterlockedPopEntrySList( ) function, 92
legacy applications, 76
InterlockedPushEntrySList( ) function, 92
libraries, 289, 290, 291, 298
Inter-Procedural Optimizations (IPO), 284,
Linux environment
285, 286
Intel MPI Library, 298
inter-process messaging, 69
OpenMP and, 164
interprocessor interrupt (IPI), 266, 268
power consumption and, 270
IntializeCriticalSection( ) function, 91 Pthreads and, 120
IntializeCriticalSectionAndSpinCount( ) live locks, 180
function, 91
LLC (Last Level Cache), 257, 263,
intra-process messaging, 69 264, 265
IP (instruction pointers), 255 load balancing
IPI (interprocessor interrupt), 266 defined, 42
IPO (Inter-Procedural Optimizations), 284, 285 loop scheduling and partitioning,
ISA (Instruction Set Architecture), 34, 244, 250 143, 144, 145, 146, 147
Itanium architecture processor affinity and, 97
multi-core for, 252, 262 VTune analyzer, 295
parallel programming problems, 207, 208, Local APIC unit, 242
209, 210 LOCK prefix, 205, 211
lock statement (C#), 117
J locked mutexes, 123
JIT (just-in-time) compilation, 33, 107 lock-holder pre-emption, 35
JMX (Java Management Extension), 68 lock-ordering deadlock, 57
Join( ) method, 112 locks
joining threads, 112, 122, 123 heavily contended, 181, 182, 183, 184,
185, 186
JVM (Java Virtual Machine), 24, 33
Intel Thread Checker, 278
live, 180
K
ordering of, 179
kernel-level threads parallel programming problems, 177
overview, 26, 27, 28, 29 synchronization primitives, 63, 64, 65,
Windows and, 75, 76, 77, 78, 79, 80, 81, 82, 66, 70
83, 84, 85, 86, 87, 88, 90, 91, 92, 93, 94, logic block, 239
95, 96, 97, 98, 99, 100 logic operators, 227
kernels loop variables, 136
defined, 27 loop-carried dependence, 138, 139,
mutexes and, 91 140, 147
semaphores and, 91 loop-independent dependence, 138
 Index 329

loops Microsoft Visual Studio, 222, 224, 225, 226,

challenges in threading, 137, 138, 139, 140, 227, 228, 230, 231, 232
141, 142, 143, 144, 145, 146, 147, Microsoft Windows. See Windows
148, 149 environment
Intel Thread Checker, 280 MIMD (multiple instruction, multiple data), 6
MISD (multiple instruction, single data), 6
M Monitor class (.NET), 118
ManualResetEvent( ) method, 117 monitors, 68
many-to-many mapping model, 28 Montecito, 262, 265
many-to-one mapping model, 28 Moore, Gordon, 7
Math Kernel Library (Intel), Moore's law, 7
289, 290, 291 MOVS operation, 207
Mauchley, John, 1 MPI (Message Passing Interface), 70, 289, 291,
MAXIMUM_PROCESSORS constant, 99 296, 297, 298, 299, 300
MAXIMUM_WAIT_OBJECTS constant, 85 MPSoC (multiprocessor systems-on-chip), 250
memory MSDN (Microsoft Developer
allocating, 202, 203 Network), 227
parallel programming problems, 190, 191, multi-core architecture
193, 194, 195, 196, 197, 198, 199, 200 background, 1, 2
threads and, 296 concurrency in software, 2, 3, 4, 5
memory barriers, 60 next steps, 271, 272
memory caching. See caching memory parallel computing platforms, 5, 6, 7, 8, 9,
10, 11, 12, 13
memory cage, 209
performance and, 14, 15, 16, 17, 18, 19
memory fences, 60
multi-core processors
messages
chip multiprocessing, 9
logging, 223, 224
hardware-based threading, 248, 249, 250,
Message Passing Interface, 70, 296, 297, 251, 252
298, 299, 300
HT Technology and, 252, 253, 254, 255,
parallel programming, 68, 69, 70 256, 257
passing, 208 interaction among, 266, 268
MFC APIs, 75, 76, 77, 78, 79, 80, 81, 82, 83, multi-threading and, 12, 13
84, 85, 86, 87, 88, 90, 91, 92, 93, 94, 95, overview, 257, 258, 259, 260, 262, 263,
96, 97, 98, 99, 100, 101, 102, 103, 104, 264, 265
105, 106
parallel threads and, 11
Michael, Maged, 190 power consumption, 268, 269, 270, 271
microprocessors multiprocessors, 260
functional units and, 239
multi-threading
parallel computing platforms, 7, 8, 9
chip, 30
Microsoft Developer Network (MSDN), 227 compiling and linking applications, 105, 106
Microsoft .NET Framework. See .NET cooperative, 29
Framework debugging applications in Windows, 224,
225, 226, 227, 228, 229, 230, 231, 232
330 Multi-Core Programming

debugging using GDB, 232, 233, 234, 235 omp_get_thread_num( ) function, 162
general debug techniques, 215, 216, 217, OMP_NUM_THREADS environment variable,
218, 219, 220, 221, 222, 223, 224 163, 166, 291
hardware and, 29, 30, 248, 249, 250, OMP_SCHEDULE environment variable, 145,
251, 252 146, 163
parallel computing platforms and, 12, 13, 19 omp_set_num_threads( ) function, 162
preemptive, 1, 28 one-to-one mapping model, 28, 29
semaphores and, 61
OOP (object-oriented programming), 37
time-slice, 8, 9
OpenMP standard
mutexes
cache optimization and, 287
data-race conditions and, 140
challenges in threading loops, 137, 138,
defined, 65, 88 139, 140, 141, 142, 143, 144, 145, 146,
priority inheritance, 183 147, 148, 149
Pthreads and, 123, 124 compilation, 164, 165
wait handles and, 118 debugging and, 165, 166, 167
Windows and, 73, 74, 85, 87, 90 environment variables, 163, 164
mutual exclusion synchronization, 54 Intel compilers, 282
Intel libraries and, 289
N Intel Thread Checker, 278, 281
Intel Thread Profiler, 296
name property (thread), 110
library functions, 162, 163
named critical sections, 157, 158
managing loops, 174
nested critical sections, 158 minimizing threading overhead, 149, 150,
.NET Framework 151, 152
creating threads, 107, 108, 109, 110 overview, 135, 136, 137
managing threads, 110, 111, 112 parallelism and, 276
parallel programming problems, 211 performance and, 144, 152, 153, 154, 155,
thread pools, 112, 113, 114, 115, 156, 157, 158, 159, 160, 161, 168, 169
116, 117 threads and, 24, 25, 27
thread synchronization, 117, 118, 119, 120 OpenMutex( ) function, 90
threading APIs, 107, 109, 110, 111, 112, OpenSemaphore( ) function, 88
113, 114, 115, 116, 117, 118, 119, 120
out-of-order execution. See ILP
non-blocking algorithms, 186, 187, 188, 189,
output dependence, 138
190, 191
Northbridge chips, 239
P
notify( ) operation, 68
notifyAll( ) operation, 68 PAL (Processor Abstraction Layer), 264
nowait clause (OpenMP), 152, 153, 155 parallel computing platforms
NPTL (Native POSIX Thread Library), 73 HT Technology, 10, 11
microprocessors, 7, 8, 9
O multi-threading on, 12, 13
overview, 5, 6, 19
omp_get_num_procs( ) function, 162 parallel for pragma (OpenMP), 139, 140, 141,
omp_get_num_threads( ) function, 162 144, 150, 157
 Index 331

parallel programming data organization recommendations,

cache-related issues, 200, 201, 202, 203, 212, 213
204, 205, 206, 207, 208, 209, 210, 211 multi-core architecture, 14, 15, 16, 17,
challenges, 42 18, 19
concurrent as term, 5 OpenMP and, 144, 152, 153, 154, 155, 156,
critical sections, 56, 57 157, 158, 159, 160, 161, 168, 169
data organization recommendations, power consumption and, 269, 270
212, 213 threading and, 41, 42
data-race conditions, 174, 176, 177 trace buffers and, 219
deadlocks, 57, 59, 177, 178, 179, 180 tuning applications, 277
designing for threads, 37, 38, 39, 40, Win32/MFC thread APIs, 76
41, 42 PGO (Profile Guided Optimizations), 284
designing with debugging in mind, $PID tracepoint keyword, 226
216, 218 pipeline pattern, 44
error diffusion algorithm, 45, 46, 47,
pipeline stalls, 211, 212
48, 49, 50, 51
PMU (Performance Memory Unit), 287
flow control-based concepts, 71, 72
heavily contended locks, 181, 182, 183, $PNAME tracepoint keyword, 226
184, 185, 186 POSIX
implementation-dependent threading, Intel Thread Checker, 278
73, 74 Native POSIX Thread Library, 73
Intel libraries and, 289, 290, 291 threading APIs, 120, 121, 122, 123, 124,
memory issues, 193, 194, 195, 196, 197, 125, 126, 127, 128, 129, 130,
198, 199, 200 131, 132
messages, 68, 69, 70 posting process, 130
non-blocking algorithms, 186, 187, 188, power consumption, 268, 269, 270, 271
189, 190, 191 preemptive multi-threading, 1, 28
parallel as term, 5 PrintCountRead( ) function, 129
patterns in, 43, 44 PrintThreads( ) function, 124
pipeline stalls, 211, 212
PrintTraceBuffer( ) method, 222, 231
synchronization, 53, 54, 55, 60, 61, 62, 63,
priority ceilings, 182
64, 65, 66, 67, 68
thread-safe functions and libraries, priority inheritance, 182, 183
192, 193 priority inversion, 181, 182, 183
too many threads, 171, 172, 173 Priority property (thread), 109
parallel sections pragma (OpenMP), private clause (OpenMP), 140, 141, 142,
140, 151 149, 152
parallel taskq pragma (OpenMP), 160, 161 processes
partitioning, loop, 143, 144, 145, 146, 147 defined, 27
patterns, parallel programming, 43, 44, 218 messaging and, 69
PAUSE instruction, 212 processor affinity
PCB (process control block), 27 defined, 28
overview, 97, 98, 100
performance
critical sections and, 157 processor order, 205
332 Multi-Core Programming

processors QR
defined, 257
depicted, 239 QueueUserWorkItem( ) function, 93, 94, 174
execution flow, 242, 243, 244
race conditions. See data-race conditions
functionality, 240
RATs (Register Allocation Tables), 256
internals, 241, 242
power consumption, 268, 269, 270, 271 read-write locks
contention and, 186
process-process messaging, 69
defined, 66
producer/consumer model
memory contention, 197, 198, 199
semaphores and, 132
recursive deadlock, 57
producer/consumer model
recursive locks, 65
synchronization and, 62, 64, 68
threads and, 40, 41 reduction clause (OpenMP), 141, 142, 147,
148, 149, 152
pthread_cond_broadcast( ) function, 129
REE (rapid execution engine), 256
pthread_cond_signal( ) function, 129
ref keyword, 120
pthread_cond_wait( ) function, 129
RegisterWaitForSingleObject( ) method,
pthread_create( ) function, 120, 121, 122
115, 116
pthread_detach( ) function, 122
relaxed consistency, 204
pthread_join( ) function, 122
release( ) operation, 63, 64
PTHREAD_MUTEX_INITIALIZER
ReleaseMutex( ) function, 90
macro, 124
ReleaseSemaphore( ) function, 88
pthread_mutex_trylock( ) function, 124
ResetEvent( ) function, 83, 84, 86, 87
pthread_mutexattr_getprotocol( )
function, 183 Resume( ) method, 112
pthread_mutexattr_setprotocol( ) ResumeThread( ) function, 80
function, 183 runnable threads, 173
Pthreads (POSIX threads) runtime scheduling scheme (OpenMP), 143,
coding, 24, 25 145, 146
creating threads, 120, 121, 122 runtime virtualization, 33
debugging, 232, 233, 234, 235
implementation-dependent S
threading, 73
Samples Over Time feature (VTune
Intel Thread Checker, 278
analyzer), 295
kernel-level threads, 27
scalability, 42, 300
managing threads, 122, 123
signaling, 125, 126, 127, 128, 129, 130 scalar processors, 244
thread pools and, 174 schedule clause (OpenMP), 143, 144, 145, 146,
thread synchronization, 123, 124 147, 150
Pthreads (POSIX threads) scheduling, loop, 143, 144, 145, 146, 147
compilation and linking, 132 sections construct (OpenMP), 151, 152, 153
signaling, 131, 132 self-deadlock, 57
pulses, 118 semaphores
defined, 88
 Index 333

Pthreads and, 130, 131, 132 SoEMT (Switch-on Event

synchronization primitives, 60, 61, 62, Multi-threading), 262
63, 68, 70 soft affinity, 97
Win32 API and, 87 software development products. See Intel
sequential consistency, 204 software development products
set scheduler-locking command, 235 Southbridge chips, 239
SetCriticalSection( ) function, 91 speedup
SetEvent( ) function, 84, 86, 87 Amdahl's law, 14, 15, 16, 17
SetMinThreads( ) method, 113 defined, 14
SetProcessorAffinityMask( ) function, 98 spin locks, 66
SetProcessPriorityBoost( ) function, 95 SSE (Streaming SIMD Extensions), 6
SetThreadName( ) function, 227, 228 SSP (Software-based Speculative
SetThreadPriorityBoost( ) function, 95 Precomputation), 287
shared clause (OpenMP), 141 start( ) method, 109
shared variables static keyword, 142, 166
debugging and, 166 static scheduling scheme (OpenMP), 143,
protecting updates, 157, 158, 159, 160 144, 145
Sieve of Eratosthenes, 195, 196, 197 stored-program model, 1
signal( ) operation, 66 STOS operation, 207
signaling strtok function (C), 193
Intel Thread Checker, 278 superscalar architecture, 244, 245, 246, 252
POSIX threads, 125, 126, 127, 128, 129, Suspend( ) method, 112
130, 131, 132 SuspendThread( ) function, 80
SIMD (single instruction, multiple data), 6 Switch-on Event Multi-threading
single-core processors (SoEMT), 262
architecture fundamentals, 237, 238, 239, SwitchToFiber( ) function, 101
240, 241, 242, 243, 244 symmetric multi-processor (SMP), 249, 283
multiprocessors with, 260 synchronization
multi-threading and, 12, 13 .NET Framework threading, 117, 118,
single-issue, single-thread (SIST) 119, 120
processor, 248 condition, 54
SISD (single instruction, single data), 6 condition variables, 66, 67, 68
SIST (single-issue, single-thread) data-race conditions and, 177
processor, 248 defined, 42
Sleep( ) method, 109, 112, 114 Itanium architecture, 207
sleep( ) system call, 234 locks, 63, 64, 65, 66
SMP (symmetric multi-processor), 249, 283 messaging and, 70
SMT (simultaneous multi-threading) mutual exclusion, 54
HT Technology and, 252 OpenMP and, 152
Itanium architecture and, 262 parallel programming constructs,
53, 54, 55
overview, 8, 30, 249
POSIX threads, 123, 124
semaphores, 60, 61, 62, 63, 68
334 Multi-Core Programming

trace buffers and, 219, 223 93, 94, 95, 96, 97, 98, 99, 100, 101, 102,
Windows threading, 87, 88, 90, 91 103, 104, 105, 106
synchronization blocks, 57 application programming models and, 32
systag (GDB), 233 applying commands to, 235
system virtualization, 33, 34, 35 boundaries of, 68
creating, 31, 32
T creating for .NET Framework, 108,
109, 110
task decomposition, 38, 39, 41, 42 critical sections and, 56, 57
task-level parallelism, 43 deadlocks, 57, 59
taskq pragma (OpenMP), 161 defined, 8, 22
Taskqueuing extension (OpenMP), 160, 161 designing for, 37, 38, 39, 40, 41, 42
flow control-based concepts, 71, 72
TDP (Thermal Design Power), 270
hardware, 29, 30, 248, 249, 250, 251, 252
TerminateThread( ) function, 80, 87
HT Technology and, 10, 11, 252, 253, 254,
testing processor affinity, 97, 98 255, 256, 257
thermal envelope, 251 implementation-dependent, 73, 74
thread affinity, 97 Intel Thread Checker, 141, 176, 277, 278,
thread command (GDB), 235 279, 280, 281
thread IDs, 227, 228 joining, 112, 122, 123
thread pools kernel-level, 26, 27, 28, 29, 75, 76, 77, 78,
.NET Framework, 112, 113, 114, 115, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88,
116, 117 90, 91, 92, 93, 94, 95, 96, 97, 98,
parallel programming problems, 174 99, 100
Windows environment, 93, 94 memory and, 296
messages, 68, 69, 70
thread priority
minimizing overhead, 149, 150, 151
.NET Framework, 109, 110
multi-core processors, 257, 258, 259, 260,
multi-threading and, 12, 13
262, 263, 264, 265, 266, 268
priority inversion, 181, 182, 183
naming, 227, 228
Pthreads and, 120
non-blocking algorithms, 186, 187, 188,
Windows environment, 95, 96
189, 190, 191
ThreadAbortException, 110 OpenMP standard, 137, 138, 139, 140,
ThreadPool class (.NET), 113, 174 141, 142, 143, 144, 145, 146, 147,
threadprivate pragma (OpenMP), 142 148, 149
threads parallel programming problems, 171,
allocating, 94 172, 173
APIs for .NET Framework, 107, 109, 110, resuming, 112
111, 112, 113, 114, 115, 116, 117, 118, suspending, 112, 172
119, 120 synchronization and, 53, 54, 55, 60, 61, 62,
APIs for POSIX, 120, 121, 122, 123, 124, 63, 64, 65, 66, 67, 68
125, 126, 127, 128, 129, 130, system view of, 22, 23, 24, 25, 26, 27, 28,
131, 132 29, 30
APIs for Windows, 75, 76, 77, 78, 79, 80, user-level, 22, 23, 24, 25, 27, 29, 100, 101,
81, 82, 83, 84, 85, 86, 87, 88, 90, 91, 92, 102, 103, 104
 Index 335

virtual environments and, 33, 34, 35 OpenMP clauses and, 142, 149
waiting on, 112 shared, 157, 158, 159, 160, 166
Threads window (Visual Studio), 225 uninitialized, 167
ThreadStart( ) function, 107, 109 virtual memory, 172, 173
$TID tracepoint keyword, 226 virtual processors, 34
time slices, 172, 173 Visual Basic language, 107
time-sharing operating systems, 1 Visual Studio (Microsoft), 222, 224, 225, 226,
time-slice multi-threading, 8, 28 227, 228, 230, 231, 232
TLP (thread-level parallelism) VMM (virtual machine monitor), 33, 34, 35
defined, 8 VMs (virtual machines), 33, 34, 35
Itanium architecture and, 262 voltage, 269, 270
multi-core processing and, 249 von Neumann, John, 1, 2
TMT (temporal multi-threading), 248, 264
$TNAME tracepoint keyword, 226 W
trace buffers, 218, 219, 220, 221, 222,
Wait( ) method, 118
223, 224
wait( ) operation, 66, 68
Trace Cache Miss event, 256
WaitAll( ) method, 117
tracepoints, 225, 226
WaitAny( ) method, 117
transistors
WaitForMultipleObjects( ) function, 84, 85
enhancing density, 272
logic blocks and, 239 WaitForSingleObject( ) function, 84, 85, 88,
90, 124
microprocessors and, 7
power consumption and, 269, 270 wavefront pattern, 44
thermal envelopes and, 251 Win32 APIs, 75, 76, 77, 78, 79, 80, 81, 82,
83, 84, 85, 86, 87, 88, 90, 91, 92, 93, 94,
TryEnterCrticalSection( ) function, 91
95, 96, 97, 98, 99, 100, 101, 102, 103, 104,
105, 106
U
Windows environment
UI (user interface), 12 atomic operations, 91, 92
UNIX platform, 120 creating threads, 76, 77, 78, 79
unlocked mutexes, 123 debugging applications, 224, 225, 226, 227,
228, 229, 230, 231, 232
user-level threads
Intel MPI Library, 298
fibers and, 100, 101, 102, 103, 104
Intel Thread Checker, 278
overview, 22, 23, 24, 25
managing threads, 80, 81
Windows support, 27, 29
multi-threaded applications, 105, 106
mutexes, 73, 74, 85, 87, 90
V
OpenMP and, 164
variables power consumption and, 270
condition, 66, 67, 68, 70, 125, 126, 127, priority inheritance, 183
128, 129 processor affinity, 97, 98, 100
environment, 163, 164 pthreads-win32 support, 120
loop, 136
336 Multi-Core Programming

thread communication, 81, 82, 83, 84, 85, thread-safe libraries, 193
86, 87 user-level threading with fibers, 100, 101,
thread pools, 93, 94 102, 103, 104
thread priority, 95, 96 work stealing method, 174
thread synchronization, 87, 88, 90, 91 work-sharing, 136, 151, 152
threading APIs, 76, 77, 78, 79, 80, 81, 82, WT_EXECUTEINIOTHREAD flag, 94
83, 84, 85, 86, 87, 88, 90, 91, 92, 93, 94,
WT_EXECUTELONGFUNCTION flag, 94
95, 96, 97, 98, 99, 100, 101, 102, 103,
104, 105, 106 WT_EXECUTIONDEFAULT flag, 94
threads and, 24, 27, 29
Continuing Education is Essential
Its a challenge we all face keeping pace with constant
change in information technology. Whether our formal
training was recent or long ago, we must all find time to
keep ourselves educated and up to date in spite of the
daily time pressures of our profession.
Intel produces technical books to help the industry
learn about the latest technologies. The focus of these publications spans the basic motivation
and origin for a technology through its practical application.

Right books, right time, from the experts

These technical books are planned to synchronize with roadmaps for technology and platforms,
in order to give the industry a head-start. They provide new insights, in an engineer-to-engineer
voice, from named experts. Sharing proven insights and design methods is intended to make it
more practical for you to embrace the latest technology with greater design freedom and re-
duced risks.
I encourage you to take full advantage of Intel Press books as a way to dive deeper into the lat-
est technologies, as you plan and develop your next generation products. They are an essential
tool for every practicing engineer or programmer. I hope you will make them a part of your con-
tinuing education tool box.

Sincerely,

Justin Rattner
Senior Fellow and Chief Technology Officer
Intel Corporation

Turn the page to learn about titles

from Intel Press for system developers
 ESSENTIAL BOOKS FOR SYSTEM DEVELOPERS

Get the most out of IA-32 platforms

The Software Optimization
Cookbook, Second Edition
High-Performance Recipes
for IA-32 Platforms
By Richard Gerber, Aart J.C. Bik, Kevin B. Smith,
Need final
and Xinmin Tian cover artwork
ISBN 0-9764832-1-1
The Software Optimization Cookbook,
Second Edition, provides updated recipes
for high-performance applications on
Intel platforms. Through simple
explanations and examples, four experts
show you how to address performance
issues with algorithms, memory access,
branch prediction, automatic
vectorization, SIMD instructions,
multiple threads, and floating-point
calculations.
Software developers learn how to take
This book simplifies the
advantage of Intel Extended Memory 64 task for engineers who
Technology (Intel EM64T), multi-core
processing, Hyper-Threading Technology, strive to develop high-

OpenMP , and multimedia extensions. This performance software...
book guides you through the growing collection
Lars Petter Endresen,
of software tools, compiler switches, and coding
Doctor of Engineering, Physics,
optimizations, showing you efficient ways to
Scandpower Petroleum Technology
improve the performance of software applications
for Intel platforms.
Highlights include:
Automatic vectorization and hints on how to guide the compiler
Compiler support for multi-threading
The performance impacts of shared L2 and L3 caches
Loop optimizations and when to use the compiler for performance gain
Use of intrinsics to exploit SIMD
Software developers who want to understand the latest techniques for delivering more perform-
ance and to fine-tune their coding skills will benefit from this book.
Programming with Intel Extended Memory 64
Technology
Migrating Software for Optimal 64-bit Performance
By Andrew Binstock
ISBN 0-9764832-0-3

A veteran technology analyst helps programmers fully

capitalize on 64-bit processing capabilities for the desktop
while ensuring full compatibility with current 32-bit
operating systems and applications. Through examples
written in C, this concise book explains how you can This book is really
enjoy the flexibility to move to 64-bit computing and
achieve better performance when working with large
practical and useful. It
datasets. thoroughly covers the
depth of the technology

Oleksiy Danikhno,
Director, Application
Development and Architecture
A4Vision, Inc.
VTune Performance Analyzer Essentials
Measurement and Tuning Techniques
for Software Developers
By James Reinders
ISBN 0-9743649-5-9

The Intel VTune Performance Analyzer illuminates

your system and everything running on it. This book is
a guide for software application developers, software
architects, quality assurance testers, and system integra-
tors who wish to use the VTune analyzer to take the
A comprehensive
guesswork out of software tuning.
approach to
increasing software
productivity
Malik S. Maxutov,
Professor and Senior Lecturer,
Moscow State Geological
Prospecting University
Intel Integrated Performance Primitives
How to Optimize Software Applications
Using Intel IPP
By Stewart Taylor
ISBN 0-9717861-3-5

The lead developer of the Intel Integrated Performance

Primitives (Intel IPP) explains how this library gives you
access to advanced processor features without having to
write processor-specific code. This introduction to Intel
IPP explores the range of possible applications, from au- Filled with
dio processing to graphics and video. Extensive examples
written in C++ show you how to solve common
comprehensive
imaging, audio/video, and graphics problems. real-world
examples...
Davis W. Frank,
Software Program Manager,
palmOne, Inc.

The Software Vectorization Handbook

Applying Multimedia Extensions
for Maximum Performance
By Aart J.C. Bik
ISBN 0-9743649-2-4

This book provides a detailed overview of compiler

optimizations that convert sequential code into a form that
exploits multimedia extensions. The primary focus is on
the C programming language and multimedia extensions
to the Intel architecture, although most conversion meth- Rarely have I
ods are easily generalized to other imperative program-
seen a book of such
ming languages and multimedia instruction sets.
a great value to
compiler writers
and application
developers alike
Robert van Engelen,
Professor,
Florida State University
Special Deals, Special Prices!
To ensure you have all the latest books
and enjoy aggressively priced discounts,
please go to this Web site:

www.intel.com/intelpress/bookbundles.htm

Bundles of our books are available,

selected especially to address the needs
of the developer. The bundles place
important complementary topics at
your fingertips, and the price for a
bundle is substantially less than
buying all the books individually.
 About Intel Press
Intel Press is the authoritative source of timely, technical books
to help software and hardware developers speed up their development
process. We collaborate only with leading industry experts to deliver
reliable, first-to-market information about the latest
technologies, processes, and strategies.

Our products are planned with the help of many people in the developer
community and we encourage you to consider becoming a customer advisor.
If you would like to help us and gain additional advance insight to the latest
technologies, we encourage you to consider the Intel Press Customer
Advisor Program. You can register here:

www.intel.com/intelpress/register.htm

For information about bulk orders or corporate sales, please send e-mail to
[email protected]

Other Developer Resources from Intel

At these Web sites you can also find valuable technical information
and resources for developers:
developer.intel.com general information for developers
www.intel.com/software content, tools, training, and the
Intel Early Access Program for
software developers
www.intel.com/software/products programming tools to help you develop
high-performance applications
www.intel.com/netcomms solutions and resources for networking
and communications
www.intel.com/technology/itj Intel Technology Journal

Intel www.intel.com/idf worldwide technical conference,

the Intel Developer Forum
PRESS
 If serial number is missing, please send an
e-mail to Intel Press at [email protected]

IMPORTANT
You can access the companion Web site for this book
on the Internet at:
www.intel.com/intelpress/mcp
Use the serial number located in the upper-right hand
corner of this page to register your book and access
additional material, including all code examples and
pointers to development resources.
Programming $69.95 USA

Multi-Core Programming
Increasing Performance through Software Multi-threading
ABOUT THE AUTHORS Akhter
Roberts
Multi-Core Programming
SHAMEEM AKHTER is a platform
Discover programming techniques for Intel multi-core architect at Intel, focusing on single
socket multi-core architecture and
Increasing Performance through Software
architecture and Hyper-Threading Technology performance analysis. He has also
worked as a senior software engineer Multi-threading

Multi-Core Programming
Software developers can no longer rely on increasing clock speeds alone with the Intel Software and Solutions
to speed up single-threaded applications; instead, to gain a competitive Group, designing application
advantage, developers must learn how to properly design their optimizations for desktop and server Shameem Akhter and Jason Roberts
applications to run in a threaded environment. Multi-core architectures platforms. Shameem holds a patent on a
have a single processor package that contains two or more processor threading interface for constraint
"execution cores," or computational engines, and deliverwith programming, developed as a part of his
appropriate softwarefully parallel execution of multiple software master's thesis in computer science.
threads. Hyper-Threading Technology enables additional threads to JASON ROBERTS is a senior software
operate on each core. engineer at Intel Corporation. Over the
past 10 years, Jason has worked on a
This book helps software developers write high-performance
number of different multi-threaded
multi-threaded code for Intel's multi-core architecture while avoiding the software products that span a wide
common parallel programming issues associated with multi-threaded range of applications targeting desktop,
programs. handheld, and embedded DSP platforms.
Highlights include:
Elements of parallel programming and multi-threading
Programming with threading APIs
OpenMP: The portable solution
Solutions to common parallel programming problems
Debugging and testing multi-threaded applications
Software development tools for multi-threading
This book is a practical, hands-on volume with immediately usable code
examples that enable readers to quickly master the necessary
programming techniques. The companion Web site contains pointers to
threading and optimization tools, code samples from the book, and
extensive technical documentation on Intel multi-core architecture

Visit our Web site at: www.intel.com/intelpress

to register your book and receive information about forthcoming books
in your area of interest

Books by Engineers, for Engineers

Printed on Recycled Material Cover Design by Ted Cyrek