EE664: Introduction to Parallel

Computing

Dr. Gaurav Trivedi

Lectures 5-14

1
 Overview
Concepts and Terminology
Parallel Computer Memory Architectures
Parallel Programming Models
Designing Parallel Programs
Parallel Algorithm Examples

2
 What is Parallel Computing? (1)

 Traditionally, software has been written for serial

computation:
– To be run on a single computer having a single Central Processing
Unit (CPU);
– A problem is broken into a discrete series of instructions.
– Instructions are executed one after another.
– Only one instruction may execute at any moment in time.

3
 What is Parallel Computing? (2)

 In the simplest sense, parallel computing is the simultaneous use of multiple

compute resources to solve a computational problem.
– To be run using multiple CPUs
– A problem is broken into discrete parts that can be solved concurrently
– Each part is further broken down to a series of instructions
 Instructions from each part execute simultaneously on different CPUs

4
 Demand for Computational Speed

There is continual demand for greater computational speed

Areas requiring great computational speed include numerical

modeling and simulation of scientific and engineering problems.

Computations must be completed within a “reasonable” time period.

5
 Parallel Computing: Resources

 The compute resources can include:

– A single computer with multiple processors;
– A single computer with (multiple) processor(s) and some
specialized computer resources (GPU, FPGA …)
– An arbitrary number of computers connected by a network;
– A combination of both.

6
 Parallel Computing: The computational problem

 The computational problem usually demonstrates

characteristics such as the ability to be:
– Broken apart into discrete pieces of work that can be solved
simultaneously;
– Execute multiple program instructions at any moment in time;
– Solved in less time with multiple compute resources than with a
single compute resource.

7
 Parallel Computing: what for? (1)

 Parallel computing is an evolution of serial computing that attempts

to emulate what has always been the state of affairs in the natural
world: many complex, interrelated events happening at the same
time, yet within a sequence.
 Some examples:
– Planetary and galactic orbits
– Weather and ocean patterns
– Tectonic plate drift
– Rush hour traffic in Guwahati/New Delhi/Bangalore/Mumbai
– Automobile assembly line
– Daily operations within a business
– Building a shopping mall

8
 Parallel Computing: what for? (2)

 Traditionally, parallel computing has been considered to

be "the high end of computing" and has been motivated
by numerical simulations of complex systems and "Grand
Challenge Problems" such as:
– weather and climate
– chemical and nuclear reactions
– biological, human genome
– geological, seismic activity
– mechanical devices - from prosthetics to spacecraft
– electronic circuits
– manufacturing processes

9
 Parallel Computing: what for? (3)

 Today, commercial applications are providing an equal or greater

driving force in the development of faster computers. These
applications require the processing of large amounts of data in
sophisticated ways. Example applications include:
– parallel databases, data mining
– oil exploration
– web search engines, web based business services
– computer-aided diagnosis in medicine
– management of national and multi-national corporations
– advanced graphics and virtual reality, particularly in the entertainment
industry
– networked video and multi-media technologies
– collaborative work environments
 Ultimately, parallel computing is an attempt to maximize the infinite
but seemingly scarce commodity called time.

10
 Example: Global Weather Forecasting
Atmosphere modeled by dividing it into 3-dimensional cells.
Calculations of each cell repeated many times to model passage of
time.

Suppose whole global atmosphere (5  108 sq.miles) divided into cells of

size 1 mile  1 mile  1 mile to a height of 10 miles (10 cells high) 
50  108 cells.

Suppose each calculation requires 2000 FLOPs (floating point

operations). In one time step, 1013 FLOPs necessary.

To forecast the weather over 7 days using 1-minute intervals, a computer

operating at 10 Gflops (1010 floating point operations/s) takes 107
seconds or over 115 days.

To perform calculation in 50 minutes requires a computer operating at 34

Tflops (34  1012 floating point operations/sec).
(IBM Blue Gene is ~500 TFLOPS) 11
 Modeling Motion of Astronomical Bodies

Each body attracted to each other body by gravitational

forces. Movement of each body predicted by calculating
total force on each body.

With N bodies, N - 1 forces to calculate for each body, or

approx. N2 calculations. (Nlog2N for an efficient approx.
algorithm). After determining new positions of bodies,
calculations repeated.

A galaxy might have 1011 stars.

If each calculation is done in 10-9 sec., it takes 1013 seconds

for one iteration using N2 algorithm (or 1013/86400 108
days)
103 sec. for one iteration using the Nlog2N algorithm.
12
 Why Parallel Computing? (1)

 This is a legitime question! Parallel computing is complex

on any aspect!

 The primary reasons for using parallel computing:

– Save time - wall clock time
– Solve larger problems
– Provide concurrency (do multiple things at the same time)

13
 Why Parallel Computing? (2)

 Other reasons might include:

– Taking advantage of non-local resources - using available compute
resources on a wide area network, or even the Internet when
local compute resources are scarce.
– Cost savings - using multiple "cheap" computing resources instead
of paying for time on a supercomputer.
– Overcoming memory constraints - single computers have very
finite memory resources. For large problems, using the memories
of multiple computers may overcome this obstacle.

14
 Limitations of Serial Computing

 Limits to serial computing - both physical and practical reasons pose

significant constraints to simply building ever faster serial computers.
 Transmission speeds - the speed of a serial computer is directly dependent
upon how fast data can move through hardware. Absolute limits are the
speed of light (30 cm/nanosecond) and the transmission limit of copper wire
(9 cm/nanosecond). Increasing speeds necessitate increasing proximity of
processing elements.
 Limits to miniaturization - processor technology is allowing an increasing
number of transistors to be placed on a chip. However, even with molecular
or atomic-level components, a limit will be reached on how small
components can be.
 Economic limitations - it is increasingly expensive to make a single processor
faster. Using a larger number of moderately fast commodity processors to
achieve the same (or better) performance is less expensive.

15
 The future

 during the past 10 years, the trends indicated by ever

faster networks, distributed systems, and multi-processor
computer architectures (even at the desktop level) clearly
show that parallelism is the future of computing.
 It will be multi-forms, mixing general purpose solutions
(your PC…) and very speciliazed solutions as IBM Cells,
ClearSpeed, GPGPU from NVidia …

16
 Who and What? (1)

 Top500.org provides statistics on parallel computing users

- the charts below are just a sample. Some things to note:
– Sectors may overlap - for example, research may be classified
research. Respondents have to choose between the two.
 "Not Specified" is by far the largest application - probably
means multiple applications.

17
Who and What? (2)

18
 Concepts and Terminology

Von Neumann Architecture

 For over 40 years, virtually all computers have followed a
common machine model known as the von Neumann
computer. Named after the Hungarian mathematician
John von Neumann.

 A von Neumann computer uses the stored-program

concept. The CPU executes a stored program that specifies
a sequence of read and write operations on the memory.

19
 Basic Design

 Basic design
– Memory is used to store both program
and data instructions
– Program instructions are coded data
which tell the computer to do something
– Data is simply information to be used by
the program
 A central processing unit (CPU) gets
instructions and/or data from
memory, decodes the instructions
and then sequentially performs them.

20
 Flynn's Classical Taxonomy

 There are different ways to classify parallel computers.

One of the more widely used classifications, in use since
1966, is called Flynn's Taxonomy.
 Flynn's taxonomy distinguishes multi-processor computer
architectures according to how they can be classified
along the two independent dimensions of Instruction and
Data. Each of these dimensions can have only one of two
possible states: Single or Multiple.

21
 Flynn Matrix

 The matrix below defines the 4 possible classifications

according to Flynn

22
 Single Instruction, Single Data (SISD)

 A serial (non-parallel) computer

 Single instruction: only one instruction
stream is being acted on by the CPU during
any one clock cycle
 Single data: only one data stream is being
used as input during any one clock cycle
 Deterministic execution
 This is the oldest and until recently, the most
prevalent form of computer
 Examples: most PCs, single CPU workstations
and mainframes

23
 Single Instruction, Multiple Data (SIMD)
 A type of parallel computer
 Single instruction: All processing units execute the same instruction at any given clock
cycle
 Multiple data: Each processing unit can operate on a different data element
 This type of machine typically has an instruction dispatcher, a very high-bandwidth
internal network, and a very large array of very small-capacity instruction units.
 Best suited for specialized problems characterized by a high degree of regularity,such as
image processing.
 Synchronous (lockstep) and deterministic execution
 Two varieties: Processor Arrays and Vector Pipelines
 Examples:
– Processor Arrays: Connection Machine CM-2, Maspar MP-1, MP-2
– Vector Pipelines: IBM 9000, Cray C90, Fujitsu VP, NEC SX-2, Hitachi S820

24
 Multiple Instruction, Single Data (MISD)
 A single data stream is fed into multiple processing units.
 Each processing unit operates on the data independently via
independent instruction streams.
 Few actual examples of this class of parallel computer have ever
existed. One is the experimental Carnegie-Mellon C.mmp computer
(1971).
 Some conceivable uses might be:
– multiple frequency filters operating on a single signal stream
 multiple cryptography algorithms attempting to crack a single coded
message.

25
 Multiple Instruction, Multiple Data (MIMD)
 Currently, the most common type of parallel computer. Most modern
computers fall into this category.
 Multiple Instruction: every processor may be executing a different
instruction stream
 Multiple Data: every processor may be working with a different data
stream
 Execution can be synchronous or asynchronous, deterministic or non-
deterministic
 Examples: most current supercomputers, networked parallel
computer "grids" and multi-processor SMP computers - including
some types of PCs.

26
 Some General Parallel Terminology

Like everything else, parallel computing has its own "jargon". Some of the
more commonly used terms associated with parallel computing are listed
below. Most of these will be discussed in more detail later.

 Task
– A logically discrete section of computational work. A task is
typically a program or program-like set of instructions that is
executed by a processor.
 Parallel Task
– A task that can be executed by multiple processors safely (yields
correct results)
 Serial Execution
– Execution of a program sequentially, one statement at a time. In
the simplest sense, this is what happens on a one processor
machine. However, virtually all parallel tasks will have sections of
a parallel program that must be executed serially.

27
 Parallel Execution
– Execution of a program by more than one task, with each task being able
to execute the same or different statement at the same moment in time.

 Shared Memory
– From a strictly hardware point of view, describes a computer architecture
where all processors have direct (usually bus based) access to common
physical memory. In a programming sense, it describes a model where
parallel tasks all have the same "picture" of memory and can directly
address and access the same logical memory locations regardless of
where the physical memory actually exists.

 Distributed Memory
– In hardware, refers to network based memory access for physical
memory that is not common. As a programming model, tasks can only
logically "see" local machine memory and must use communications to
access memory on other machines where other tasks are executing.

28
 Communications
– Parallel tasks typically need to exchange data. There are several ways this
can be accomplished, such as through a shared memory bus or over a
network, however the actual event of data exchange is commonly
referred to as communications regardless of the method employed.
 Synchronization
– The coordination of parallel tasks in real time, very often associated with
communications. Often implemented by establishing a synchronization
point within an application where a task may not proceed further until
another task(s) reaches the same or logically equivalent point.
– Synchronization usually involves waiting by at least one task, and can
therefore cause a parallel application's wall clock execution time to
increase.

29
 Granularity
– In parallel computing, granularity is a qualitative measure of the ratio of
computation to communication.
– Coarse: relatively large amounts of computational work are done
between communication events
– Fine: relatively small amounts of computational work are done between
communication events
 Observed Speedup
– Observed speedup of a code which has been parallelized, defined as:
wall-clock time of serial execution
wall-clock time of parallel execution
– One of the simplest and most widely used indicators for a parallel
program's performance.

30
 Parallel Overhead
– The amount of time required to coordinate parallel tasks, as opposed to
doing useful work. Parallel overhead can include factors such as:
 Task start-up time
 Synchronizations
 Data communications
 Software overhead imposed by parallel compilers, libraries, tools, operating
system, etc.
 Task termination time
 Massively Parallel
– Refers to the hardware that comprises a given parallel system - having
many processors. The meaning of many keeps increasing, but currently
BG/L pushes this number to 6 digits.

31
 Scalability
– Refers to a parallel system's (hardware and/or software) ability to
demonstrate a proportionate increase in parallel speedup with
the addition of more processors. Factors that contribute to
scalability include:
 Hardware - particularly memory-cpu bandwidths and network
communications
 Application algorithm
 Parallel overhead related
 Characteristics of your specific application and coding

32
 Parallel Computing

Using more than one computer, or a computer with more than

one processor, to solve a problem.

Motives
n computers operating simultaneously can achieve the result n
times faster - it will not be n times faster for various reasons.

Other motives include: fault tolerance, larger amount of memory

available, ...

33
 Speedup Factor

Execution time using one processor (best sequential algorithm) ts

S(p) = = t
Execution time using a multiprocessor with p processors p

where ts is execution time on a single processor and tp is execution time on a

multiprocessor.

S(p) : increase in speed by using multiprocessor.

Use best sequential algorithm with single processor system. Underlying

algorithm for parallel implementation might be (and is usually) different.

Speedup factor can also be cast in terms of computational steps:

Number of computational steps using one processor

S(p) = Number of parallel computational steps with p processors
34
 **here Maximum Speedup

Maximum speedup is usually p with p processors (linear speedup).

Possible to get superlinear speedup (greater than p) but usually there

is a specific reason such as:

• Extra memory in multiprocessor system

• Nondeterministic algorithm

35
 Maximum Speedup
Amdahl’s law
ts
fts (1-f)ts

Serial section Parallelizable sections

(a) One processor ...

(b) Multiple
processors
...

p processors

tp (1-f)ts / p
36
 Speedup factor is given by:

ts 1 1
S ( p)   lim S ( p) 
(1  f )t s (1  f ) p  f
ft s  f
p p

This equation is known as Amdahl’s law.

Even with infinite number of processors, maximum

speedup is limited to 1/f.

Example: With only 5% of computation being serial, maximum speedup is 20,

irrespective of number of processors.

37
 Superlinear Speedup
Example - searching

(a) Searching each sub-space sequentially

Start Time

ts
ts/p

Sub-space Dt
search

xts/p
Solution found

x indeterminate

38
(b) Searching each sub-space in parallel

Speed-up then given by

ts
x´ +Dt
p
S(p) =
Dt
Dt

Solution found
39
 Worst case for sequential search when solution found in last sub-
space search. Then parallel version offers greatest benefit, i.e.

p-1
p ´ ts + Dt
S(p) =   as Dt tends to go to zero
Dt

Least advantage for parallel version when solution found in first sub-
space search of the sequential search, i.e.

Dt
S(p) = =1
Dt
Actual speed-up depends upon which subspace holds solution but could
be extremely large.
40
 Conventional Computer
Consists of a processor executing a program stored in a (main)
memory:

Main memory

Instructions (to processor)

Data (to or from processor)

Processor

Addresses start at 0 and extend to 2b - 1

41
Parallel Computers
Shared memory vs. Distributed memory

Shared Memory Multiprocessor System

Multiple processors connected to multiple memory modules such
that each processor can access any memory module :

Memory module
One
address
space

Interconnection
network

42
Processors
 Quad Pentium Shared Memory Multiprocessor

Processor Processor Processor Processor

L1 cache L1 cache L1 cache L1 cache
L2 Cache L2 Cache L2 Cache L2 Cache

Bus interface Bus interface Bus interface Bus interface

Processor/
memory
bus

I/O interface Memory controller

I/O bus

Memory
Shared memory
43
Need to address Cache coherency problem!
 Shared Memory Multiprocessors

Threads - programmer decomposes program into individual parallel

sequences, (threads), each being able to access shared variables.

Sequential programming language with preprocessor compiler

directives to declare shared variables and specify parallelism.

Example: OpenMP - industry standard - needs OpenMP compiler

44
 Message-Passing Multicomputer

Complete computers connected through an interconnection network:

Interconnection
network
Messages
Processor

Local
memory

Computers

45
Interconnection Networks

2- and 3-dimensional meshes

Hypercube

Trees

Using Switches:
Crossbar
Multistage interconnection networks

46
One-dimensional array
Links Computer/Processor

Two-dimensional array (mesh)

Links

47
 Ring

Two-dimensional Torus

48
 Three-dimensional hypercube

110 111

100 101

010 011

000 001

49
 Four-dimensional hypercube

0110 0111 1110 1111

0100 0101 1100 1101

0010 0011 1010 1011

0000 0001 1000 1001

Hypercubes were very popular in 1980’s

50
 Tree

Root

Switch
Links element

Processors

51
 Crossbar switch

Memories

Processors Switches

52
 Multistage Interconnection Network
Example: Omega network

2 x 2 switch elements
(straight-through or
Inputs crossover connections) Outputs

000 000
001 001

010 010
011 011

100 100
101 101

110 110
111 111
53
 Embedding a ring onto a hypercube

110 111

100 101

010 011

000 001

000 001 011 010 110 111 101 100

What is this sequence called? 3-bit Gray Code 54

 Embedding a 2-D Mesh onto a hypercube

11 11 0110

01 01 0110 01 0111

00
0000 0001 0011 0010 0110 0111 0101 0100 1100 1101

2-bit graycode

4-bit graycode
55
 Distributed Shared Memory

Making main memory of group of interconnected computers look as

though a single memory with single address space. Then can use
shared memory programming techniques.

Interconnection
network
Messages
Processor

Shared
memory

Computers

56
 Networked Computers as a Computing Platform

A network of computers became a very attractive alternative to

expensive supercomputers and parallel computer systems for high-
performance computing in early 1990’s.

Notable early projects:

Berkeley NOW (network of workstations).
NASA Beowulf project.

Key Advantages:
 Very high performance workstations and PCs readily available at low cost.

 The latest processors can be easily incorporated into the system as they
become available.

 Existing software can be used or modified.

57
 Parallel Algorithm Examples:
Odd-Even Transposition Sort
Basic idea is bubble sort, but concurrently comparing
odd indexed elements with an adjacent element, then
even indexed elements.

If there are n elements in an array and there are n/2

processors. The algorithm is effectively O(n)!

58
 Parallel Algorithm Examples:
Odd Even Transposition Sort
Initial array: Worst case scenario.
 6, 5, 4, 3, 2, 1, 0
6, 4, 5, 2, 3, 0, 1 Phase 1
4, 6, 2, 5, 0, 3, 1 Phase 2
4, 2, 6, 0, 5, 1, 3 Phase 1
2, 4, 0, 6, 1, 5, 3 Phase 2
2, 0, 4, 1, 6, 3, 5 Phase 1
0, 2, 1, 4, 3, 6, 5 Phase 2
0, 1, 2, 3, 4, 5, 6 Phase 1

Only 7 passes are required 59

 Memory architectures

 Shared Memory
 Distributed Memory
 Hybrid Distributed-Shared Memory
 Memory architectures

 Shared Memory
 Distributed Memory
 Hybrid Distributed-Shared Memory
 Shared Memory : UMA vs. NUMA

 Uniform Memory Access (UMA):

– Most commonly represented today by Symmetric Multiprocessor (SMP)
machines
– Identical processors
– Equal access and access times to memory
– Sometimes called CC-UMA - Cache Coherent UMA. Cache coherent
means if one processor updates a location in shared memory, all the
other processors know about the update. Cache coherency is
accomplished at the hardware level.
 Non-Uniform Memory Access (NUMA):
– Often made by physically linking two or more SMPs
– One SMP can directly access memory of another SMP
– Not all processors have equal access time to all memories
– Memory access across link is slower
– If cache coherency is maintained, then may also be called CC-NUMA -
Cache Coherent NUMA
 Shared Memory: Pro and Con

 Advantages
– Global address space provides a user-friendly programming perspective
to memory
– Data sharing between tasks is both fast and uniform due to the proximity
of memory to CPUs
 Disadvantages:
– Primary disadvantage is the lack of scalability between memory and
CPUs. Adding more CPUs can geometrically increases traffic on the
shared memory-CPU path, and for cache coherent systems, geometrically
increase traffic associated with cache/memory management.
– Programmer responsibility for synchronization constructs that insure
"correct" access of global memory.
– Expense: it becomes increasingly difficult and expensive to design and
produce shared memory machines with ever increasing numbers of
processors.
 Distributed Memory
 Like shared memory systems, distributed memory systems vary widely but share a
common characteristic. Distributed memory systems require a communication network
to connect inter-processor memory.
 Processors have their own local memory. Memory addresses in one processor do not
map to another processor, so there is no concept of global address space across all
processors.
 Because each processor has its own local memory, it operates independently. Changes
it makes to its local memory have no effect on the memory of other processors. Hence,
the concept of cache coherency does not apply.
 When a processor needs access to data in another processor, it is usually the task of the
programmer to explicitly define how and when data is communicated. Synchronization
between tasks is likewise the programmer's responsibility.
 The network "fabric" used for data transfer varies widely, though it can can be as simple
as Ethernet.
 Distributed Memory: Pro and Con

 Advantages
– Memory is scalable with number of processors. Increase the number of
processors and the size of memory increases proportionately.
– Each processor can rapidly access its own memory without interference
and without the overhead incurred with trying to maintain cache
coherency.
– Cost effectiveness: can use commodity, off-the-shelf processors and
networking.
 Disadvantages
– The programmer is responsible for many of the details associated with
data communication between processors.
– It may be difficult to map existing data structures, based on global
memory, to this memory organization.
 Hybrid Distributed-Shared Memory
 The largest and fastest computers in the world today employ both shared and
distributed memory architectures.

 The shared memory component is usually a cache coherent SMP machine.

Processors on a given SMP can address that machine's memory as global.
 The distributed memory component is the networking of multiple SMPs.
SMPs know only about their own memory - not the memory on another SMP.
Therefore, network communications are required to move data from one SMP
to another.
 Current trends seem to indicate that this type of memory architecture will
continue to prevail and increase at the high end of computing for the
foreseeable future.
 Advantages and Disadvantages: whatever is common to both shared and
distributed memory architectures.
 Hybrid Distributed-Shared Memory
Summarizing a few of the key characteristics of shared and
distributed memory machines
 Parallel Programming Models

 Overview
 Shared Memory Model
 Threads Model
 Message Passing Model
 Data Parallel Model
 Other Models
 Overview

 There are several parallel programming models in

common use:
– Shared Memory
– Threads
– Message Passing
– Data Parallel
– Hybrid
 Parallel programming models exist as an abstraction above
hardware and memory architectures.
 Overview

 Although it might not seem apparent, these models are NOT specific to a
particular type of machine or memory architecture. In fact, any of these
models can (theoretically) be implemented on any underlying hardware.
 Shared memory model on a distributed memory machine: Kendall Square
Research (KSR) ALLCACHE approach.
– Machine memory was physically distributed, but appeared to the user as a single
shared memory (global address space). Generically, this approach is referred to as
"virtual shared memory".
– Note: although KSR is no longer in business, there is no reason to suggest that a
similar implementation will not be made available by another vendor in the future.
– Message passing model on a shared memory machine: MPI on SGI Origin.
 The SGI Origin employed the CC-NUMA type of shared memory architecture,
where every task has direct access to global memory. However, the ability to
send and receive messages with MPI, as is commonly done over a network of
distributed memory machines, is not only implemented but is very commonly
used.
 Overview

 Which model to use is often a combination of what is

available and personal choice. There is no "best" model,
although there certainly are better implementations of
some models over others.
 The following sections describe each of the models
mentioned above, and also discuss some of their actual
implementations.
 Shared Memory Model

 In the shared-memory programming model, tasks share a common

address space, which they read and write asynchronously.
 Various mechanisms such as locks / semaphores may be used to
control access to the shared memory.
 An advantage of this model from the programmer's point of view is
that the notion of data "ownership" is lacking, so there is no need to
specify explicitly the communication of data between tasks. Program
development can often be simplified.
 An important disadvantage in terms of performance is that it
becomes more difficult to understand and manage data locality.
 Shared Memory Model: Implementations

 On shared memory platforms, the native compilers

translate user program variables into actual memory
addresses, which are global.
 No common distributed memory platform
implementations currently exist. However, as mentioned
previously in the Overview section, the KSR ALLCACHE
approach provided a shared memory view of data even
though the physical memory of the machine was
distributed.
 Threads Model

 In the threads model of parallel programming, a single process can have multiple,
concurrent execution paths.
 Perhaps the most simple analogy that can be used to describe threads is the concept of
a single program that includes a number of subroutines:
– The main program a.out is scheduled to run by the native operating system. a.out loads and
acquires all of the necessary system and user resources to run.
– a.out performs some serial work, and then creates a number of tasks (threads) that can be
scheduled and run by the operating system concurrently.
– Each thread has local data, but also, shares the entire resources of a.out. This saves the
overhead associated with replicating a program's resources for each thread. Each thread also
benefits from a global memory view because it shares the memory space of a.out.
– A thread's work may best be described as a subroutine within the main program. Any thread
can execute any subroutine at the same time as other threads.
– Threads communicate with each other through global memory (updating address locations).
This requires synchronization constructs to insure that more than one thread is not updating
the same global address at any time.
– Threads can come and go, but a.out remains present to provide the necessary shared
resources until the application has completed.
 Threads are commonly associated with shared memory architectures and operating
systems.
 Threads Model Implementations

 From a programming perspective, threads implementations commonly comprise:

– A library of subroutines that are called from within parallel source code
– A set of compiler directives imbedded in either serial or parallel source code
 In both cases, the programmer is responsible for determining all parallelism.
 Threaded implementations are not new in computing. Historically, hardware vendors
have implemented their own proprietary versions of threads. These implementations
differed substantially from each other making it difficult for programmers to develop
portable threaded applications.
 Unrelated standardization efforts have resulted in two very different implementations
of threads: POSIX Threads and OpenMP.
 POSIX Threads
– Library based; requires parallel coding
– Specified by the IEEE POSIX 1003.1c standard (1995).
– C Language only
– Commonly referred to as Pthreads.
– Most hardware vendors now offer Pthreads in addition to their proprietary threads
implementations.
– Very explicit parallelism; requires significant programmer attention to detail.
 Threads Model: OpenMP

 OpenMP
– Compiler directive based; can use serial code
– Jointly defined and endorsed by a group of major computer hardware
and software vendors. The OpenMP Fortran API was released October 28,
1997. The C/C++ API was released in late 1998.
– Portable / multi-platform, including Unix and Windows NT platforms
– Available in C/C++ and Fortran implementations
– Can be very easy and simple to use - provides for "incremental
parallelism"
 Microsoft has its own implementation for threads, which is not
related to the UNIX POSIX standard or OpenMP.
 Message Passing Model

 The message passing model demonstrates the following

characteristics:
– A set of tasks that use their own local memory during
computation. Multiple tasks can reside on the same physical
machine as well across an arbitrary number of machines.
– Tasks exchange data through communications by sending and
receiving messages.
– Data transfer usually requires cooperative operations to be
performed by each process. For example, a send operation must
have a matching receive operation.
 Message Passing Model Implementations: MPI

 From a programming perspective, message passing implementations

commonly comprise a library of subroutines that are imbedded in
source code. The programmer is responsible for determining all
parallelism.
 Historically, a variety of message passing libraries have been available
since the 1980s. These implementations differed substantially from
each other making it difficult for programmers to develop portable
applications.
 In 1992, the MPI Forum was formed with the primary goal of
establishing a standard interface for message passing
implementations.
 Part 1 of the Message Passing Interface (MPI) was released in 1994.
Part 2 (MPI-2) was released in 1996. Both MPI specifications are
available on the web at
www.mcs.anl.gov/Projects/mpi/standard.html.
 Message Passing Model Implementations: MPI
 MPI is now the "de facto" industry standard for message passing, replacing
virtually all other message passing implementations used for production
work. Most, if not all of the popular parallel computing platforms offer at
least one implementation of MPI. A few offer a full implementation of MPI-2.
 For shared memory architectures, MPI implementations usually don't use a
network for task communications. Instead, they use shared memory (memory
copies) for performance reasons.
 Data Parallel Model

 The data parallel model demonstrates the following characteristics:

– Most of the parallel work focuses on performing operations on a data set.
The data set is typically organized into a common structure, such as an
array or cube.
– A set of tasks work collectively on the same data structure, however, each
task works on a different partition of the same data structure.
– Tasks perform the same operation on their partition of work, for example,
"add 4 to every array element".
 On shared memory architectures, all tasks may have access to the
data structure through global memory. On distributed memory
architectures the data structure is split up and resides as "chunks" in
the local memory of each task.
 Data Parallel Model Implementations

 Programming with the data parallel model is usually accomplished by

writing a program with data parallel constructs. The constructs can be
called to a data parallel subroutine library or, compiler directives
recognized by a data parallel compiler.
 Fortran 90 and 95 (F90, F95): ISO/ANSI standard extensions to Fortran
77.
– Contains everything that is in Fortran 77
– New source code format; additions to character set
– Additions to program structure and commands
– Variable additions - methods and arguments
– Pointers and dynamic memory allocation added
– Array processing (arrays treated as objects) added
– Recursive and new intrinsic functions added
– Many other new features
 Implementations are available for most common parallel platforms.
 Data Parallel Model Implementations
 High Performance Fortran (HPF): Extensions to Fortran 90 to support
data parallel programming.
– Contains everything in Fortran 90
– Directives to tell compiler how to distribute data added
– Assertions that can improve optimization of generated code added
– Data parallel constructs added (now part of Fortran 95)
– Implementations are available for most common parallel platforms.
 Compiler Directives: Allow the programmer to specify the distribution
and alignment of data. Fortran implementations are available for
most common parallel platforms.
 Distributed memory implementations of this model usually have the
compiler convert the program into standard code with calls to a
message passing library (MPI usually) to distribute the data to all the
processes. All message passing is done invisibly to the programmer.
 Other Models

 Other parallel programming models besides those previously

mentioned certainly exist, and will continue to evolve along with the
ever changing world of computer hardware and software.
 Only three of the more common ones are mentioned here.
– Hybrid
– Single Program Multiple Data
– Multiple Program Multiple Data
 Hybrid

 In this model, any two or more parallel programming models are

combined.
 Currently, a common example of a hybrid model is the combination of
the message passing model (MPI) with either the threads model
(POSIX threads) or the shared memory model (OpenMP). This hybrid
model lends itself well to the increasingly common hardware
environment of networked SMP machines.
 Another common example of a hybrid model is combining data
parallel with message passing. As mentioned in the data parallel
model section previously, data parallel implementations (F90, HPF) on
distributed memory architectures actually use message passing to
transmit data between tasks, transparently to the programmer.
 Single Program Multiple Data (SPMD)

 Single Program Multiple Data (SPMD):

 SPMD is actually a "high level" programming model that can be built
upon any combination of the previously mentioned parallel
programming models.
 A single program is executed by all tasks simultaneously.
 At any moment in time, tasks can be executing the same or different
instructions within the same program.
 SPMD programs usually have the necessary logic programmed into
them to allow different tasks to branch or conditionally execute only
those parts of the program they are designed to execute. That is,
tasks do not necessarily have to execute the entire program - perhaps
only a portion of it.
 All tasks may use different data
 Multiple Program Multiple Data (MPMD)

 Multiple Program Multiple Data (MPMD):

 Like SPMD, MPMD is actually a "high level" programming
model that can be built upon any combination of the
previously mentioned parallel programming models.
 MPMD applications typically have multiple executable
object files (programs). While the application is being run
in parallel, each task can be executing the same or
different program as other tasks.
 All tasks may use different data
 Some Important Points

 Automatic vs. Manual Parallelization

 Understand the Problem and the Program
 Partitioning
 Communications
 Synchronization
 Data Dependencies
 Load Balancing
 Granularity
 I/O
 Limits and Costs of Parallel Programming
 Performance Analysis and Tuning
 Automatic vs. Manual Parallelization
 Understand the Problem and the Program
 Partitioning
 Communications
 Synchronization
 Data Dependencies
 Load Balancing
 Granularity
 I/O
 Limits and Costs of Parallel Programming
 Performance Analysis and Tuning
 Designing and developing parallel programs has characteristically
been a very manual process. The programmer is typically responsible
for both identifying and actually implementing parallelism.
 Very often, manually developing parallel codes is a time consuming,
complex, error-prone and iterative process.
 For a number of years now, various tools have been available to assist
the programmer with converting serial programs into parallel
programs. The most common type of tool used to automatically
parallelize a serial program is a parallelizing compiler or pre-processor.
 A parallelizing compiler generally works in two different ways:
– Fully Automatic
 The compiler analyzes the source code and identifies opportunities for
parallelism.
 The analysis includes identifying inhibitors to parallelism and possibly a cost
weighting on whether or not the parallelism would actually improve
performance.
 Loops (do, for) loops are the most frequent target for automatic
parallelization.
– Programmer Directed
 Using "compiler directives" or possibly compiler flags, the programmer
explicitly tells the compiler how to parallelize the code.
 May be able to be used in conjunction with some degree of automatic
parallelization also.
 If you are beginning with an existing serial code and have time or
budget constraints, then automatic parallelization may be the answer.
However, there are several important caveats that apply to automatic
parallelization:
– Wrong results may be produced
– Performance may actually degrade
– Much less flexible than manual parallelization
– Limited to a subset (mostly loops) of code
– May actually not parallelize code if the analysis suggests there are
inhibitors or the code is too complex
– Most automatic parallelization tools are for Fortran
 The remainder of this section applies to the manual method of
developing parallel codes.
 Automatic vs. Manual Parallelization
 Understand the Problem and the Program
 Partitioning
 Communications
 Synchronization
 Data Dependencies
 Load Balancing
 Granularity
 I/O
 Limits and Costs of Parallel Programming
 Performance Analysis and Tuning
 Undoubtedly, the first step in developing parallel software
is to first understand the problem that you wish to solve in
parallel. If you are starting with a serial program, this
necessitates understanding the existing code also.
 Before spending time in an attempt to develop a parallel
solution for a problem, determine whether or not the
problem is one that can actually be parallelized.
Example of Parallelizable Problem
Calculate the potential energy for each of several thousand
independent conformations of a molecule. When done,
find the minimum energy conformation.

 This problem is able to be solved in parallel. Each of the

molecular conformations is independently determinable.
The calculation of the minimum energy conformation is
also a parallelizable problem.
Example of a Non-parallelizable Problem
Calculation of the Fibonacci series (1,1,2,3,5,8,13,21,...) by
use of the formula:
F(k + 2) = F(k + 1) + F(k)

 This is a non-parallelizable problem because the

calculation of the Fibonacci sequence as shown would
entail dependent calculations rather than independent
ones. The calculation of the k + 2 value uses those of both
k + 1 and k. These three terms cannot be calculated
independently and therefore, not in parallel.
 Identify the program's hotspots

 Know where most of the real work is being done. The

majority of scientific and technical programs usually
accomplish most of their work in a few places.
 Profilers and performance analysis tools can help here
 Focus on parallelizing the hotspots and ignore those
sections of the program that account for little CPU usage.
 Identify bottlenecks in the program

 Are there areas that are disproportionately slow, or cause

parallelizable work to halt or be deferred? For example,
I/O is usually something that slows a program down.
 May be possible to restructure the program or use a
different algorithm to reduce or eliminate unnecessary
slow areas
 Other considerations

 Identify inhibitors to parallelism. One common class of

inhibitor is data dependence, as demonstrated by the
Fibonacci sequence above.
 Investigate other algorithms if possible. This may be the
single most important consideration when designing a
parallel application.
 Automatic vs. Manual Parallelization
 Understand the Problem and the Program
 Partitioning
 Communications
 Synchronization
 Data Dependencies
 Load Balancing
 Granularity
 I/O
 Limits and Costs of Parallel Programming
 Performance Analysis and Tuning
 One of the first steps in designing a parallel program is to
break the problem into discrete "chunks" of work that can
be distributed to multiple tasks. This is known as
decomposition or partitioning.
 There are two basic ways to partition computational work
among parallel tasks:
– domain decomposition
and
– functional decomposition
 Domain Decomposition

 In this type of partitioning, the data associated with a

problem is decomposed. Each parallel task then works on
a portion of of the data.
 Partitioning Data

 There are different ways to partition data

 Functional Decomposition
 In this approach, the focus is on the computation that is to be
performed rather than on the data manipulated by the computation.
The problem is decomposed according to the work that must be
done. Each task then performs a portion of the overall work.
 Functional decomposition lends itself well to problems that can be
split into different tasks. For example
– Ecosystem Modeling
– Signal Processing
– Climate Modeling
 Ecosystem Modeling

 Each program calculates the population of a given group,

where each group's growth depends on that of its
neighbors. As time progresses, each process calculates its
current state, then exchanges information with the
neighbor populations. All tasks then progress to calculate
the state at the next time step.
 Signal Processing

 An audio signal data set is passed through four distinct computational

filters. Each filter is a separate process. The first segment of data must
pass through the first filter before progressing to the second. When it
does, the second segment of data passes through the first filter. By
the time the fourth segment of data is in the first filter, all four tasks
are busy.
 Climate Modeling
 Each model component can be thought of as a separate task. Arrows
represent exchanges of data between components during
computation: the atmosphere model generates wind velocity data
that are used by the ocean model, the ocean model generates sea
surface temperature data that are used by the atmosphere model,
and so on.

 Combining these two types of problem decomposition is common and

natural.
 Automatic vs. Manual Parallelization
 Understand the Problem and the Program
 Partitioning
 Communications
 Synchronization
 Data Dependencies
 Load Balancing
 Granularity
 I/O
 Limits and Costs of Parallel Programming
 Performance Analysis and Tuning
 Who Needs Communications?

 The need for communications between tasks depends upon your problem
 You DON'T need communications
– Some types of problems can be decomposed and executed in parallel with virtually
no need for tasks to share data. For example, imagine an image processing
operation where every pixel in a black and white image needs to have its color
reversed. The image data can easily be distributed to multiple tasks that then act
independently of each other to do their portion of the work.
– These types of problems are often called embarrassingly parallel because they are
so straight-forward. Very little inter-task communication is required.
 You DO need communications
– Most parallel applications are not quite so simple, and do require tasks to share
data with each other. For example, a 3-D heat diffusion problem requires a task to
know the temperatures calculated by the tasks that have neighboring data.
Changes to neighboring data has a direct effect on that task's data.
 Factors to Consider (1)

 There are a number of important factors to consider when

designing your program's inter-task communications
 Cost of communications
– Inter-task communication virtually always implies overhead.
– Machine cycles and resources that could be used for computation
are instead used to package and transmit data.
– Communications frequently require some type of synchronization
between tasks, which can result in tasks spending time "waiting"
instead of doing work.
– Competing communication traffic can saturate the available
network bandwidth, further aggravating performance problems.
 Factors to Consider (2)

 Latency vs. Bandwidth

– latency is the time it takes to send a minimal (0 byte) message
from point A to point B. Commonly expressed as microseconds.
– bandwidth is the amount of data that can be communicated per
unit of time. Commonly expressed as megabytes/sec.
– Sending many small messages can cause latency to dominate
communication overheads. Often it is more efficient to package
small messages into a larger message, thus increasing the
effective communications bandwidth.
 Factors to Consider (3)

 Visibility of communications
– With the Message Passing Model, communications are explicit
and generally quite visible and under the control of the
programmer.
– With the Data Parallel Model, communications often occur
transparently to the programmer, particularly on distributed
memory architectures. The programmer may not even be able to
know exactly how inter-task communications are being
accomplished.
 Factors to Consider (4)

 Synchronous vs. asynchronous communications

– Synchronous communications require some type of "handshaking"
between tasks that are sharing data. This can be explicitly structured in
code by the programmer, or it may happen at a lower level unknown to
the programmer.
– Synchronous communications are often referred to as blocking
communications since other work must wait until the communications
have completed.
– Asynchronous communications allow tasks to transfer data
independently from one another. For example, task 1 can prepare and
send a message to task 2, and then immediately begin doing other work.
When task 2 actually receives the data doesn't matter.
– Asynchronous communications are often referred to as non-blocking
communications since other work can be done while the communications
are taking place.
– Interleaving computation with communication is the single greatest
benefit for using asynchronous communications.
 Factors to Consider (5)

 Scope of communications
– Knowing which tasks must communicate with each other is critical
during the design stage of a parallel code. Both of the two
scopings described below can be implemented synchronously or
asynchronously.
– Point-to-point - involves two tasks with one task acting as the
sender/producer of data, and the other acting as the
receiver/consumer.
– Collective - involves data sharing between more than two tasks,
which are often specified as being members in a common group,
or collective.
 Collective Communications

 Examples
 Factors to Consider (6)

 Efficiency of communications
– Very often, the programmer will have a choice with regard to
factors that can affect communications performance. Only a few
are mentioned here.
– Which implementation for a given model should be used? Using
the Message Passing Model as an example, one MPI
implementation may be faster on a given hardware platform than
another.
– What type of communication operations should be used? As
mentioned previously, asynchronous communication operations
can improve overall program performance.
– Network media - some platforms may offer more than one
network for communications. Which one is best?
 Factors to Consider (7)

 Overhead and Complexity

 Automatic vs. Manual Parallelization
 Understand the Problem and the Program
 Partitioning
 Communications
 Synchronization
 Data Dependencies
 Load Balancing
 Granularity
 I/O
 Limits and Costs of Parallel Programming
 Performance Analysis and Tuning
 Types of Synchronization

 Barrier
– Usually implies that all tasks are involved
– Each task performs its work until it reaches the barrier. It then stops, or "blocks".
– When the last task reaches the barrier, all tasks are synchronized.
– What happens from here varies. Often, a serial section of work must be done. In other cases,
the tasks are automatically released to continue their work.
 Lock / semaphore
– Can involve any number of tasks
– Typically used to serialize (protect) access to global data or a section of code. Only one task at
a time may use (own) the lock / semaphore / flag.
– The first task to acquire the lock "sets" it. This task can then safely (serially) access the
protected data or code.
– Other tasks can attempt to acquire the lock but must wait until the task that owns the lock
releases it.
– Can be blocking or non-blocking
 Synchronous communication operations
– Involves only those tasks executing a communication operation
– When a task performs a communication operation, some form of coordination is required with
the other task(s) participating in the communication. For example, before a task can perform a
send operation, it must first receive an acknowledgment from the receiving task that it is OK to
send.
– Discussed previously in the Communications section.
 Automatic vs. Manual Parallelization
 Understand the Problem and the Program
 Partitioning
 Communications
 Synchronization
 Data Dependencies
 Load Balancing
 Granularity
 I/O
 Limits and Costs of Parallel Programming
 Performance Analysis and Tuning
 Definitions

 A dependence exists between program statements when

the order of statement execution affects the results of the
program.
 A data dependence results from multiple use of the same
location(s) in storage by different tasks.
 Dependencies are important to parallel programming
because they are one of the primary inhibitors to
parallelism.
 Examples (1): Loop carried data dependence

DO 500 J = MYSTART,MYEND
A(J) = A(J-1) * 2.0500
CONTINUE

 The value of A(J-1) must be computed before the value of

A(J), therefore A(J) exhibits a data dependency on A(J-1).
Parallelism is inhibited.
 If Task 2 has A(J) and task 1 has A(J-1), computing the
correct value of A(J) necessitates:
– Distributed memory architecture - task 2 must obtain the value of
A(J-1) from task 1 after task 1 finishes its computation
– Shared memory architecture - task 2 must read A(J-1) after task 1
updates it
 Examples (2): Loop independent data dependence

task 1 task 2
------ ------
X = 2 X = 4
. .
. .
Y = X**2 Y = X**3

 As with the previous example, parallelism is inhibited. The value of Y

is dependent on:
– Distributed memory architecture - if or when the value of X is
communicated between the tasks.
– Shared memory architecture - which task last stores the value of X.
 Although all data dependencies are important to identify when
designing parallel programs, loop carried dependencies are
particularly important since loops are possibly the most common
target of parallelization efforts.
 How to Handle Data Dependencies?

 Distributed memory architectures - communicate required

data at synchronization points.
 Shared memory architectures -synchronize read/write
operations between tasks.
 Automatic vs. Manual Parallelization
 Understand the Problem and the Program
 Partitioning
 Communications
 Synchronization
 Data Dependencies
 Load Balancing
 Granularity
 I/O
 Limits and Costs of Parallel Programming
 Performance Analysis and Tuning
 Definition

 Load balancing refers to the practice of distributing work among tasks

so that all tasks are kept busy all of the time. It can be considered a
minimization of task idle time.
 Load balancing is important to parallel programs for performance
reasons. For example, if all tasks are subject to a barrier
synchronization point, the slowest task will determine the overall
performance.
 How to Achieve Load Balance? (1)

 Equally partition the work each task receives

– For array/matrix operations where each task performs similar
work, evenly distribute the data set among the tasks.
– For loop iterations where the work done in each iteration is
similar, evenly distribute the iterations across the tasks.
– If a heterogeneous mix of machines with varying performance
characteristics are being used, be sure to use some type of
performance analysis tool to detect any load imbalances. Adjust
work accordingly.
 How to Achieve Load Balance? (2)

 Use dynamic work assignment

– Certain classes of problems result in load imbalances even if data is
evenly distributed among tasks:
 Sparse arrays - some tasks will have actual data to work on while others have
mostly "zeros".
 Adaptive grid methods - some tasks may need to refine their mesh while
others don't.
 N-body simulations - where some particles may migrate to/from their original
task domain to another task's; where the particles owned by some tasks
require more work than those owned by other tasks.
– When the amount of work each task will perform is intentionally variable,
or is unable to be predicted, it may be helpful to use a scheduler - task
pool approach. As each task finishes its work, it queues to get a new
piece of work.
– It may become necessary to design an algorithm which detects and
handles load imbalances as they occur dynamically within the code.
 Automatic vs. Manual Parallelization
 Understand the Problem and the Program
 Partitioning
 Communications
 Synchronization
 Data Dependencies
 Load Balancing
 Granularity
 I/O
 Limits and Costs of Parallel Programming
 Performance Analysis and Tuning
 Definitions

 Computation / Communication Ratio:

– In parallel computing, granularity is a qualitative measure of the
ratio of computation to communication.
– Periods of computation are typically separated from periods of
communication by synchronization events.
 Fine grain parallelism
 Coarse grain parallelism
 Fine-grain Parallelism

 Relatively small amounts of computational work are

done between communication events
 Low computation to communication ratio
 Facilitates load balancing
 Implies high communication overhead and less
opportunity for performance enhancement
 If granularity is too fine it is possible that the
overhead required for communications and
synchronization between tasks takes longer than
the computation.
 Coarse-grain Parallelism

 Relatively large amounts of computational

work are done between
communication/synchronization events
 High computation to communication ratio
 Implies more opportunity for performance
increase
 Harder to load balance efficiently
 Which is Best?

 The most efficient granularity is dependent on the

algorithm and the hardware environment in which it runs.
 In most cases the overhead associated with
communications and synchronization is high relative to
execution speed so it is advantageous to have coarse
granularity.
 Fine-grain parallelism can help reduce overheads due to
load imbalance.
 Automatic vs. Manual Parallelization
 Understand the Problem and the Program
 Partitioning
 Communications
 Synchronization
 Data Dependencies
 Load Balancing
 Granularity
 I/O
 Limits and Costs of Parallel Programming
 Performance Analysis and Tuning
 The bad News

 I/O operations are generally regarded as inhibitors to

parallelism
 Parallel I/O systems are immature or not available for all
platforms
 In an environment where all tasks see the same filespace,
write operations will result in file overwriting
 Read operations will be affected by the fileserver's ability
to handle multiple read requests at the same time
 I/O that must be conducted over the network (NFS, non-
local) can cause severe bottlenecks
 The good News

 Some parallel file systems are available. For example:

– GPFS: General Parallel File System for AIX (IBM)
– Lustre: for Linux clusters (Cluster File Systems, Inc.)
– PVFS/PVFS2: Parallel Virtual File System for Linux clusters
(Clemson/Argonne/Ohio State/others)
– PanFS: Panasas ActiveScale File System for Linux clusters (Panasas, Inc.)
– HP SFS: HP StorageWorks Scalable File Share. Lustre based parallel file
system (Global File System for Linux) product from HP
 The parallel I/O programming interface specification for MPI has been
available since 1996 as part of MPI-2. Vendor and "free"
implementations are now commonly available.
 Some Options

 If you have access to a parallel file system, investigate using it. If you
don't, keep reading...
 Rule #1: Reduce overall I/O as much as possible
 Confine I/O to specific serial portions of the job, and then use parallel
communications to distribute data to parallel tasks. For example, Task
1 could read an input file and then communicate required data to
other tasks. Likewise, Task 1 could perform write operation after
receiving required data from all other tasks.
 For distributed memory systems with shared filespace, perform I/O in
local, non-shared filespace. For example, each processor may have
/tmp filespace which can used. This is usually much more efficient
than performing I/O over the network to one's home directory.
 Create unique filenames for each tasks' input/output file(s)
 Automatic vs. Manual Parallelization
 Understand the Problem and the Program
 Partitioning
 Communications
 Synchronization
 Data Dependencies
 Load Balancing
 Granularity
 I/O
 Limits and Costs of Parallel Programming
 Performance Analysis and Tuning
 Amdahl's Law

Amdahl's Law states that potential

program speedup is defined by the
fraction of code (P) that can be
parallelized:
1
speedup = --------
1 - P

 If none of the code can be parallelized, P = 0 and the

speedup = 1 (no speedup). If all of the code is parallelized,
P = 1 and the speedup is infinite (in theory).
 If 50% of the code can be parallelized, maximum speedup
= 2, meaning the code will run twice as fast.
 Amdahl's Law

 Introducing the number of processors performing the

parallel fraction of work, the relationship can be modeled
by
1
speedup = ------------
P + S
---
N

 where P = parallel fraction, N = number of processors and

S = serial fraction
 Amdahl's Law

 It soon becomes obvious that there are limits to the

scalability of parallelism. For example, at P = .50, .90
and .99 (50%, 90% and 99% of the code is parallelizable)

speedup
--------------------------------
N P = .50 P = .90 P = .99
----- ------- ------- -------
10 1.82 5.26 9.17
100 1.98 9.17 50.25
1000 1.99 9.91 90.99
10000 1.99 9.91 99.02
 Amdahl's Law

 However, certain problems demonstrate increased performance by

increasing the problem size. For example:
– 2D Grid Calculations 85 seconds 85%
– Serial fraction 15 seconds 15%
 We can increase the problem size by doubling the grid dimensions
and halving the time step. This results in four times the number of
grid points and twice the number of time steps. The timings then look
like:
– 2D Grid Calculations 680 seconds 97.84%
– Serial fraction 15 seconds 2.16%
 Problems that increase the percentage of parallel time with their size
are more scalable than problems with a fixed percentage of parallel
time.
 Complexity

 In general, parallel applications are much more complex than

corresponding serial applications, perhaps an order of magnitude. Not
only do you have multiple instruction streams executing at the same
time, but you also have data flowing between them.
 The costs of complexity are measured in programmer time in virtually
every aspect of the software development cycle:
– Design
– Coding
– Debugging
– Tuning
– Maintenance
 Adhering to "good" software development practices is essential when
when working with parallel applications - especially if somebody
besides you will have to work with the software.
 Portability

 Thanks to standardization in several APIs, such as MPI, POSIX threads,

HPF and OpenMP, portability issues with parallel programs are not as
serious as in years past. However...
 All of the usual portability issues associated with serial programs
apply to parallel programs. For example, if you use vendor
"enhancements" to Fortran, C or C++, portability will be a problem.
 Even though standards exist for several APIs, implementations will
differ in a number of details, sometimes to the point of requiring code
modifications in order to effect portability.
 Operating systems can play a key role in code portability issues.
 Hardware architectures are characteristically highly variable and can
affect portability.
 Resource Requirements

 The primary intent of parallel programming is to decrease execution

wall clock time, however in order to accomplish this, more CPU time
is required. For example, a parallel code that runs in 1 hour on 8
processors actually uses 8 hours of CPU time.
 The amount of memory required can be greater for parallel codes
than serial codes, due to the need to replicate data and for overheads
associated with parallel support libraries and subsystems.
 For short running parallel programs, there can actually be a decrease
in performance compared to a similar serial implementation. The
overhead costs associated with setting up the parallel environment,
task creation, communications and task termination can comprise a
significant portion of the total execution time for short runs.
 Scalability

 The ability of a parallel program's performance to scale is a result of a

number of interrelated factors. Simply adding more machines is rarely
the answer.
 The algorithm may have inherent limits to scalability. At some point,
adding more resources causes performance to decrease. Most parallel
solutions demonstrate this characteristic at some point.
 Hardware factors play a significant role in scalability. Examples:
– Memory-cpu bus bandwidth on an SMP machine
– Communications network bandwidth
– Amount of memory available on any given machine or set of machines
– Processor clock speed
 Parallel support libraries and subsystems software can limit scalability
independent of your application.
 Automatic vs. Manual Parallelization
 Understand the Problem and the Program
 Partitioning
 Communications
 Synchronization
 Data Dependencies
 Load Balancing
 Granularity
 I/O
 Limits and Costs of Parallel Programming
 Performance Analysis and Tuning
 As with debugging, monitoring and analyzing parallel
program execution is significantly more of a challenge
than for serial programs.
 A number of parallel tools for execution monitoring and
program analysis are available.
 Some are quite useful; some are cross-platform also.
 Work remains to be done, particularly in the area of
scalability.
 Parallel Examples

 Array Processing
 PI Calculation
 Simple Heat Equation
 1-D Wave Equation
 Array Processing

 This example demonstrates calculations on 2-dimensional array

elements, with the computation on each array element being
independent from other array elements.
 The serial program calculates one element at a time in sequential
order.
 Serial code could be of the form:
do j = 1,n
do i = 1,n
a(i,j) = fcn(i,j)
end do
end do
 The calculation of elements is independent of one another - leads to
an embarrassingly parallel situation.
 The problem should be computationally intensive.
 Array Processing Solution 1

 Arrays elements are distributed so that each processor owns a portion of an array
(subarray).
 Independent calculation of array elements insures there is no need for communication
between tasks.
 Distribution scheme is chosen by other criteria, e.g. unit stride (stride of 1) through the
subarrays. Unit stride maximizes cache/memory usage.
 Since it is desirable to have unit stride through the subarrays, the choice of a
distribution scheme depends on the programming language.
 After the array is distributed, each task executes the portion of the loop corresponding
to the data it owns. For example, with Fortran block distribution:

do j = mystart, myend
do i = 1,n
a(i,j) = fcn(i,j)
end do
end do

 Notice that only the outer loop variables are different from the serial solution.
 Array Processing Solution 1
One possible implementation

 Implement as SPMD model.

 Master process initializes array, sends info to worker
processes and receives results.
 Worker process receives info, performs its share of
computation and sends results to master.
 Using the Fortran storage scheme, perform block
distribution of the array.
 Pseudo code solution: red highlights changes for
parallelism.
Array Processing Solution 1
One possible implementation
 Array Processing Solution 2: Pool of Tasks

 The previous array solution demonstrated static load

balancing:
– Each task has a fixed amount of work to do
– May be significant idle time for faster or more lightly loaded
processors - slowest tasks determines overall performance.
 Static load balancing is not usually a major concern if all
tasks are performing the same amount of work on
identical machines.
 If you have a load balance problem (some tasks work
faster than others), you may benefit by using a "pool of
tasks" scheme.
 Array Processing Solution 2
Pool of Tasks Scheme

 Two processes are employed

 Master Process:
– Holds pool of tasks for worker processes to do
– Sends worker a task when requested
– Collects results from workers
 Worker Process: repeatedly does the following
– Gets task from master process
– Performs computation
– Sends results to master
 Worker processes do not know before runtime which portion of array
they will handle or how many tasks they will perform.
 Dynamic load balancing occurs at run time: the faster tasks will get
more work to do.
 Pseudo code solution: red highlights changes for parallelism.
Array Processing Solution 2 Pool of Tasks Scheme
 Pi Calculation

 The value of PI can be calculated in a number of ways.

Consider the following method of approximating PI
– Inscribe a circle in a square
– Randomly generate points in the square
– Determine the number of points in the square that are also in the
circle
– Let r be the number of points in the circle divided by the number
of points in the square
– PI ~ 4 r
 Note that the more points generated, the better the
approximation
 Discussion

 In the above pool of tasks example, each task calculated

an individual array element as a job. The computation to
communication ratio is finely granular.
 Finely granular solutions incur more communication
overhead in order to reduce task idle time.
 A more optimal solution might be to distribute more work
with each job. The "right" amount of work is problem
dependent.
 Algorithm

npoints = 10000
circle_count = 0
do j = 1,npoints
generate 2 random numbers between
0 and 1
xcoordinate = random1 ;
ycoordinate = random2
if (xcoordinate, ycoordinate)
inside circle then circle_count =
circle_count + 1
end do
PI = 4.0*circle_count/npoints

Note that most of the time in running this program

would be spent executing the loop
Leads to an embarrassingly parallel solution
- Computationally intensive
- Minimal communication
- Minimal I/O
 PI Calculation
Parallel Solution

 Parallel strategy: break the loop into

portions that can be executed by the
tasks.
 For the task of approximating PI:
– Each task executes its portion of the
loop a number of times.
– Each task can do its work without
requiring any information from the
other tasks (there are no data
dependencies).
– Uses the SPMD model. One task acts as
master and collects the results.
 Pseudo code solution: red highlights
changes for parallelism.
PI Calculation
Parallel Solution
 Simple Heat Equation

 Most problems in parallel computing require communication among

the tasks. A number of common problems require communication
with "neighbor" tasks.
 The heat equation describes the temperature
change over time, given initial temperature
distribution and boundary conditions.
 A finite differencing scheme is employed to
solve the heat equation numerically on a
square region.
 The initial temperature is zero on the
boundaries and high in the middle.
 The boundary temperature is held at zero.
 For the fully explicit problem, a time stepping
algorithm is used. The elements of a 2-dimensional array represent
the temperature at points on the square.
 Simple Heat Equation

 The calculation of an element is dependent upon neighbor

element values.

 A serial program would contain code like:

 Parallel Solution 1

 Implement as an SPMD model

 The entire array is partitioned and distributed
as subarrays to all tasks. Each task owns a
portion of the total array.
 Determine data dependencies
– interior elements belonging to a task are independent of other tasks
– border elements are dependent upon a neighbor task's data,
necessitating communication.
 Master process sends initial info to workers, checks for convergence
and collects results
 Worker process calculates solution, communicating as necessary with
neighbor processes
 Pseudo code solution: red highlights changes for parallelism.
Parallel Solution 1
 Parallel Solution 2
Overlapping Communication and Computation

 In the previous solution, it was assumed that blocking

communications were used by the worker tasks. Blocking
communications wait for the communication process to complete
before continuing to the next program instruction.
 In the previous solution, neighbor tasks communicated border data,
then each process updated its portion of the array.
 Computing times can often be reduced by using non-blocking
communication. Non-blocking communications allow work to be
performed while communication is in progress.
 Each task could update the interior of its part of the solution array
while the communication of border data is occurring, and update its
border after communication has completed.
 Pseudo code for the second solution: red highlights changes for non-
blocking communications.
Parallel Solution 2
Overlapping Communication and Computation
 1-D Wave Equation

 In this example, the amplitude along a uniform, vibrating

string is calculated after a specified amount of time has
elapsed.
 The calculation involves:
– the amplitude on the y axis
– i as the position index along the x axis
– node points imposed along the string
– update of the amplitude at discrete time steps.
 1-D Wave Equation

 The equation to be solved is the one-dimensional wave

equation:

where c is a constant

 Note that amplitude will depend on previous timesteps (t,

t-1) and neighboring points (i-1, i+1). Data dependence
will mean that a parallel solution will involve
communications.
 1-D Wave Equation
Parallel Solution
 Implement as an SPMD model
 The entire amplitude array is partitioned and distributed as subarrays
to all tasks. Each task owns a portion of the total array.
 Load balancing: all points require equal work, so the points should be
divided equally
 A block decomposition would have the work partitioned into the
number of tasks as chunks, allowing each task to own mostly
contiguous data points.
 Communication need only occur on data borders. The larger the block
size the less the communication.
1-D Wave Equation Parallel Solution