Grace: Safe Multithreaded Programming for C/C++

Emery D. Berger

Ting Yang

Tongping Liu

Gene Novark

Dept. of Computer Science

University of Massachusetts, Amherst
Amherst, MA 01003
{emery,tingy,tonyliu,gnovark}@cs.umass.edu

Abstract
The shift from single to multiple core architectures means
that programmers must write concurrent, multithreaded programs in order to increase application performance. Unfortunately, multithreaded applications are susceptible to numerous errors, including deadlocks, race conditions, atomicity
violations, and order violations. These errors are notoriously
difficult for programmers to debug.
This paper presents Grace, a software-only runtime system that eliminates concurrency errors for a class of multithreaded programs: those based on fork-join parallelism.
By turning threads into processes, leveraging virtual memory protection, and imposing a sequential commit protocol, Grace provides programmers with the appearance of
deterministic, sequential execution, while taking advantage
of available processing cores to run code concurrently and
efficiently. Experimental results demonstrate Graces effectiveness: with modest code changes across a suite of
computationally-intensive benchmarks (116 lines), Grace
can achieve high scalability and performance while preventing concurrency errors.
Categories and Subject Descriptors D.1.3 [Software]:
Concurrent ProgrammingParallel Programming; D.2.0
[Software Engineering]: Protection mechanisms
General Terms Performance, Reliability
Keywords Concurrency, determinism, deterministic concurrency, fork-join, sequential semantics

1.

Introduction

While the past two decades have seen dramatic increases

in processing power, the problems of heat dissipation and

Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. To copy otherwise, to republish, to post on servers or to redistribute
to lists, requires prior specific permission and/or a fee.
OOPSLA 2009, October 2529, 2009, Orlando, Florida, USA.
c 2009 ACM 978-1-60558-734-9/09/10. . . $10.00
Copyright

energy consumption now limit the ability of hardware manufacturers to speed up chips by increasing their clock rate.
This phenomenon has led to a major shift in computer architecture, where single-core CPUs have been replaced by
CPUs consisting of a number of processing cores.
The implication of this switch is that the performance of
sequential applications is no longer increasing with each new
generation of processors, because the individual processing
components are not getting faster. On the other hand, applications rewritten to use multiple threads can take advantage
of these available computing resources to increase their performance by executing their computations in parallel across
multiple CPUs.
Unfortunately, writing multithreaded programs is challenging. Concurrent multithreaded applications are susceptible to a wide range of errors that are notoriously difficult
to debug [29]. For example, multithreaded programs that fail
to employ a canonical locking order can deadlock [16]. Because the interleavings of threads are non-deterministic, programs that do not properly lock shared data structures can
suffer from race conditions [30]. A related problem is atomicity violations, where programs may lock and unlock individual objects but fail to ensure the atomicity of multiple
object updates [14]. Another class of concurrency errors is
order violations, where a program depends on a sequence of
threads that the scheduler may not provide [26].
This paper introduces Grace, a runtime system that eliminates concurrency errors for a particular class of multithreaded programs: those that employ fully-structured, or
fork-join based parallelism to increase performance.
While fork-join parallelism does not capture all possible parallel programs, it is a popular model of parallel program execution: systems based primarily on forkjoin parallelism include Cilk, Intels Threading Building
Blocks [35], OpenMP, and the fork-join framework proposed for Java [24]. Perhaps the most prominent use of forkjoin parallelism today is in Googles Map-Reduce framework, a library that is used to implement a number of Google
services [9, 34]. However, none of these prevent concurrency

errors, which are difficult even for expert programmers to

avoid [13].
Grace manages the execution of multithreaded programs
with fork-join parallelism so that they become behaviorally
equivalent to their sequential counterparts: every thread
spawn becomes a sequential function invocation, and locks
become no-ops.
This execution model eliminates most concurrency errors
that can arise due to multithreading (see Table 1). By converting lock operations to no-ops, Grace eliminates deadlocks. By committing state changes deterministically, Grace
eliminates race conditions. By executing threads in program
order, Grace eliminates atomicity violations and greatly reduces the risk of order violations. Finally, by enforcing sequential semantics and thus sequential consistency, Grace
eliminates the need for programmers to reason about complex underlying memory models.
To exploit available computing resources (multiple CPUs
or cores), Grace employs a combination of speculative
thread execution, together with a sequential commit protocol
that ensures sequential semantics. By replacing threads with
processes and providing appropriate shared memory mappings, Grace leverages process isolation, page protection
and virtual memory mappings to provide isolation and full
support for speculative execution on conventional hardware.
Under Grace, threads execute optimistically, writing their
updates speculatively but locally. As long as the threads do
not conflict, that is, they do not have read-write dependencies
on the same memory location, then Grace can safely commit
their effects. In case of a conflict, Grace commits the earliest
thread in program order from the conflicting set of threads.
Rather than executing threads atomically, Grace uses events
like thread spawns and joins as commit points that divide
execution into pieces of work, and enforces a deterministic
execution that matches a sequential execution.
This deterministic execution model allows programmers
to reason about their programs as if they were serial programs, making them easier to understand and debug [2].
Traditionally, when programmers reorganize thread interactions to obtain reasonable performance (e.g., by selecting an
appropriate grain size, reducing contention, and minimizing
the size of critical sections), they run risk of introducing new,
difficult-to-debug concurrency errors. Grace not only lifts
the burden of using locks or atomic sections on programmers, but also allows them to optimize performance without
the risk of compromising correctness.
We evaluate Graces performance on a suite of CPUintensive, fork-join based multithreaded applications, as well
as a microbenchmark we designed to explore the space of
programs for which Grace will be most effective. We also
evaluate Graces ability to avoid a selection of concurrency
bugs taken from the literature. Experimental results show
that Grace ensures the correct execution of otherwise-buggy
concurrent code. While Grace does not guarantee concur-

// Run f(x) and g(y) in parallel.

t1 = spawn f(x);
t2 = spawn g(y);
// Wait for both to complete.
sync;

Figure 1. A multithreaded program (using Cilk syntax for

clarity).
// Run f(x) to completion, then g(y).
t1 = spawn f(x);
t2 = spawn g(y);
// Wait for both to complete.
sync;

Figure 2. Its sequential counterpart (elided operations

struck out).
rency for unchanged programs, we found that minor changes
(116 lines of source code) were enough to allow Grace to
achieve comparable scalability and performance to the standard (unsafe) threads library across most of our benchmark
suite, while ensuring safe execution.
The remainder of this paper is organized as follows. Section 2 outlines the sequential semantics that Grace provides.
Section 3 describes the software mechanisms that Grace uses
to enable speculative execution with low overhead. Section 4
presents the commit protocol that enforces sequential semantics, and explains how Grace can support I/O together
with optimistic concurrency. Section 5 describes our experimental methodology. Section 6 then presents experimental results across a benchmark suite of concurrent, multithreaded computation kernels, a microbenchmark that explores Graces performance characteristics, and a suite of
concurrency errors. Section 7 surveys related work, Section 8 describes future directions, and Section 9 concludes.

2.

Sequential Semantics

To illustrate the effect of running Grace, we use the example

shown in Figure 1, which for clarity uses Cilk-style thread
operations rather than the subset of the pthreads API that
Grace supports. Here, spawn creates a thread to execute the
argument function, and sync waits for all threads spawned
in the current scope to complete.
This example program executes the two functions f and g
asynchronously (as threads), and waits for them to complete.
If f and g share state, this execution could result in atomicity
violations or race conditions; if these functions acquire locks
in different orders, then they could deadlock. Now consider
the version of this program shown in Figure 2, where calls
to spawn and sync (struck out) are ignored.
The second program is the serial elision [5] of the first
all parallel function calls have been elided. The result is a
serial program that, by definition, cannot suffer from concur-

Concurrency Error
Deadlock
Race condition
Atomicity violation
Order violation

Cause
cyclic lock acquisition
unguarded updates
unguarded, interleaved updates
threads scheduled in unexpected order

Prevention by Grace
locks converted to no-ops
all updates committed deterministically
threads run atomically
threads execute in program order

Table 1. The concurrency errors that Grace addresses, their causes, and how Grace eliminates them.
rency errors. Because the executions of f(x) and g(y) are
not interleaved and execute deterministically, atomicity violations or race conditions are impossible. Similarly, the ordering of execution of these functions is fixed, so there cannot be order violations. Finally, a sequential program does
not need locks, so eliding them prevents deadlock.
2.1

Programming Model

Grace enforces deterministic execution of programs that

rely on fully structured or fork-join parallelism, such as
master-slave parallelism or parallelized divide-and-conquer,
where each division step forks off children threads and waits
for them to complete. These programs have a straightforward
sequential counterpart: the serial elision described above.
For convenience, Grace exports its operations as a subset of
the popular POSIX pthreads API, although it does not
support the full range of pthreads semantics.
Graces current target class of applications is applications running fork-join style, CPU-intensive operations. At
present, Grace is not suitable for reactive programs like
server applications, and does not support programs with
concurrency control through synchronization primitives like
condition variables, or other programs that are inherently
concurrent: that is, their serial elision does not result in a
program that exhibits the same semantics.
Note that while Grace is able to prevent a number of concurrency errors, it cannot eliminate errors that are external
to the program itself. For example, Grace does not attempt
to detect or prevent errors like file system deadlocks (e.g.,
through flock()) or due to message-passing dependencies
on distributed systems.

3.

Support for Speculation

Grace achieves concurrent speedup of multithreaded programs by executing threads speculatively, then committing
their updates in program order (see Section 4). A key challenge is how to enable low-overhead thread speculation in
C/C++.
One possible candidate would be some form of transactional memory [17, 36]. Unfortunately, no existing or proposed transactional memory system provides all of the features that Grace requires:
full compatibility with C and C++ and commodity hard-

ware,
full support for long-lived transactions,

complete isolation of updates from other threads, i.e.,

strong atomicity [6],

support for irrevocable actions including I/O and memory

management, and
extremely low runtime and space overhead.

Existing software transactional memory (STM) systems

are optimized for short transactions, generally demarcated
with atomic clauses. These systems do not effectively
support long-lived transactions, which either abort whenever conflicting shorter-lived transactions commit their state
first, or must switch to single-threaded mode to ensure fair
progress. They also often preclude the use of irrevocable
actions (e.g., I/O) inside transactions [40].
Most importantly, STMs typically incur substantial space
and runtime overhead (around 3X) for fully-isolated memory updates inside transactions. While compiler optimizations can reduce this cost on unshared data [37], transactions
must still incur this overhead on shared data.
In the absence of sophisticated compiler analyses, we
found that the overheads of conventional log-based STMs
are unacceptable for the long transactions that Grace targets.
We attempted to employ Suns state-of-the-art TL2 STM
system [11] using Pin [28] to instrument reads and writes
that call the appropriate TL2 function (transactional reads
and writes). Unlike most programs using TL2 (including
the STAMP transaction benchmark suite), the transactions
here comprise every read and write. In all of our tests, the
length of the logs becomes excessive, causing TL2 to run
out of memory.
To meet its requirements, Grace employs a novel virtualmemory based software transactional memory with a number of key features. First, it supports fully-isolated threads
of arbitrary length (in terms of the number of memory addresses read or written). Second, its performance overhead
is amortized over the length of a threads execution rather
than being incurred on every access, so that threads that
run for more than a few milliseconds effectively run at
full speed. Third, it supports threads with arbitrary operations, including irrevocable I/O calls (see Section 4). Finally,
Grace works with existing C and C++ applications running
on commodity hardware.
3.1

Processes as Threads

Our key insight is that we can implement efficient software

transactional memory by treating threads as processes: in-

committed (shared) pages & version numbers

uncommitted (private) pages

thread
begin

reads

writes

{}

{}

{1}

{}

{1,4}

{}

{1,4}

{4}

protected
read-only
unprotected
(copy-on-write)

thread
end

Figure 3. An overview of execution in Grace. Processes emulate threads (Section 3.1) with private mappings to mmapped files
that hold committed pages and version numbers for globals and the heap (Sections 3.2 and 3.3). Threads run concurrently but
are committed in sequential order: each thread waits until its logical predecessor has terminated in order to preserve sequential
semantics (Section 4). Grace then compares the version numbers of the read pages to the committed versions. If they match,
Grace commits the writes and increments version numbers; otherwise, it discards the pages and rolls back.
stead of spawning new threads, Grace forks off new processes. Because each thread is in fact a separate process,
it is possible to use standard memory protection functions
and signal handlers to track reads and writes to memory.
Grace tracks accesses to memory at a page granularity, trading imprecision of object tracking for speed. Crucially, because only the first read or write to each page needs to be
tracked, all subsequent operations proceed at full speed.
To create the illusion that these processes are executing
in a shared address space, Grace uses memory mapped files
to share the heap and globals across processes. Each process has two mappings to the heap and globals: a shared
mapping that reflects the latest committed state, and a local (per-process), copy-on-write mapping that each process
uses directly. In addition, Grace establishes a shared and local map of an array of version numbers. Grace uses these
version numbersone for each page in the heap and global
areato decide when it is safe to commit updates.
3.2

Globals

Grace uses a fixed-size file to hold the globals, which it locates in the program image through linker-defined variables.
In ELF executables, the symbol end indicates the first address after uninitialized global data. Grace uses an ld-based
linker script to identify the area that indicates the start of
the global data. In addition, this linker script instructs the
linker to page align and separate read-only and global areas
of memory. This separation reduces the risk of false sharing
by ensuring that writes to a global object never conflict with
reads of read-only data.
3.3

Heap Organization

Grace also uses a fixed-size mapping (currently 512MB) to

hold the heap. It embeds the heap data structure into the beginning of the memory-mapped file itself. This organization

elegantly solves the problem of rolling back memory allocations. Grace rolls back memory allocations just as it rolls
back any other updates to heap data. Any conflict causes the
heap to revert to an earlier version.
However, a nave implementation of the allocator would
give rise to an unacceptably large number of conflicts: any
threads that perform memory allocations would conflict. For
example, consider a basic freelist-based allocator. Any allocation or deallocation updates a freelist pointer. Thus, any
time two threads both invoke malloc or free on the samesized object, one thread will be forced to roll back because
both threads are updating the page holding that pointer.
To avoid this problem of inadvertent rollbacks, Grace
uses a scalable per-thread heap organization that is loosely
based on Hoard [3] and built with Heap Layers [4]. Grace
divides the heap into a fixed number of sub-heaps (currently
16). Each thread uses a hash of its process id to obtain the
index of the heap it uses for all memory operations (malloc
and free).
This isolation of each threads memory operations from
the others allows threads to operate independently most
of the time. Each sub-heap is initially seeded with a pagealigned 64K chunk of memory. As long as a thread does
not exhaust its own sub-heaps pool of memory, it will operate independently from any other sub-heap. If it runs out
of memory, it obtains another 64K chunk from the global
allocator. This allocation only causes a conflict with another
thread if that thread also runs out of memory during the same
period of time.
This allocation strategy has two benefits. First, it minimizes the number of false conflicts created by allocations
from the main heap. Second, it avoids an important source
of false sharing. Because each thread uses different pages to
satisfy object allocation requests, objects allocated by one
thread are unlikely to be on the same pages as objects al-

located by another thread (except when both threads hash to

the same sub-heap). This heap organization ensures that conflicts only arise when allocated memory from a parent thread
is passed to children threads, or when objects allocated by
one thread are then accessed by another, later thread.
To further reduce false sharing, Graces heap rounds up
large object requests (8K or larger) to a multiple of the
system page size (4K), ensuring that large objects never
overlap, regardless of which thread allocated them.
3.4

void atomicBegin (void) {

// Roll back to here on abort.
// Saves PC, registers, stack.
context.commit();
// Reset pages seen (for signal handler).
pages.clear();
// Reset global and heap protection.
globals.begin();
heap.begin();
}

Thread Execution

Figure 3 presents an overview of Graces execution of a

thread. This example is simplified: recall that Grace does not
always execute entire threads atomically. Atomic execution
begins at program startup (main()), and whenever a new
thread is spawned. It ends (is committed) not only when a
thread ends, but also when a thread spawns a child or joins
(syncs) a previously-spawned child thread.
Before the program begins, Grace establishes shared and
local mappings for the heap and globals. It also establishes
the mappings for the version numbers associated with each
page in both the heap and global area. Because these pages
are zero-filled on-demand, this mapping implicitly initializes
the version numbers to zero. A pages version number is
incremented only on a successful commit, so it is equivalent
to its total number of successful commits to date.
Initialization
Grace initializes state tracking at the beginning of program execution and at the start of every thread by invoking
atomicBegin (Figure 4). Grace first saves the execution
context (program counter, registers, and stack contents) and
sets the protection of every page to PROT NONE, so that any
access triggers a fault. It also clears both its read and write
sets, which hold the addresses of every page read or written.
Execution
Grace tracks accesses to pages by handling SEGV protection faults. The first access to each page is treated as a read.
Grace adds the page address to the read set, and then sets
the protection for the page to read-only. If the application
later writes to the page, Grace adds the page to the write
set, and then removes all protection from the page. Thus,
in the worst case, a thread incurs two minor page faults for
every page that it visits. While protection faults and signals
are expensive, their cost is quickly amortized even for relatively short-lived threads (e.g., a millisecond or more), as
Section 6.2 shows.

Figure 4. Pseudo-code for atomic begin.

checks to see whether the read set is empty, at which point

it can safely commit. While this situation may appear to
be unlikely, it is common when multiple threads are being
created inside a for loop, and thus the application is only
reading local variables from registers. Allowing commits in
this case is an important optimization, because otherwise,
Grace would have to pause the thread until its immediate
predecessorthe last thread it has spawnedhas committed. As Section 4 explains, this step is required to provide
sequential semantics.
Committing
Once a thread has finished executing and any logically preceding threads have already completed, Grace establishes
locks on all files holding memory mappings using interprocess mutexes (in the call to lock()) and proceeds to
check whether it is safe to commit its updates. Notice that
this serialization only occurs during commits; thread execution is entirely concurrent.
Grace first performs a consistency check, comparing the
version numbers for every page in the read set against the
committed versions both for the heap and the globals. If
they all match, it is safe for Grace to commit the writes,
which it does by copying the contents of each page into the
corresponding page in the shared images. It then relinquishes
the file locks and resumes execution.
If, however, any of the version numbers do not match,
Grace invokes atomicAbort to abort the current execution (Figure 5). Grace issues a madvise(MADV DONTNEED)
call to discard any updates to the heap and globals, which
forces all new accesses to use memory from the shared
(committed) pages. It then unlocks the file maps and reexecutes, copying the saved stack over the current stack and
then jumping into the previously saved execution context.

Completion
At the end of each atomically-executed regionthe end
of main() or an individual thread, right before a thread
spawn, and right before joining another threadGrace invokes atomicEnd (Figure 5), which attempts to commit
all updates by calling atomicCommit (Figure 6). It first

4.

Sequential Commit

Grace provides strong isolation of threads, ensuring that they

do not interfere with each other when executing speculatively. However, this isolation on its own does not guarantee
sequential semantics because it does not prescribe any order.

void atomicEnd (void) {

if (!atomicCommit())
atomicAbort();
}
void atomicAbort (void) {
// Throw away changes.
heap.abort();
globals.abort();
// Jump back to saved context.
context.abort();
}

Figure 5. Pseudo-code for atomic end and abort.

bool
//
//
//
if

atomicCommit (void) {
If havent read or written anything,
we dont have to wait or commit;
update local view of memory & return.
(heap.nop() && globals.nop()) {
heap.updateAll();
globals.updateAll();
return true;

}
// Wait for immediate predecessor
// to complete.
waitExited(predecessor);
// Now try to commit state. Iff we succeed,
// return true.
// Lock to make check & commit atomic.
lock();
bool committed = false;
// Ensure heap and globals consistent.
if (heap.consistent() &&
globals.consistent()) {
// OK, all consistent: commit.
heap.commit();
globals.commit();
xio.commit(); // commits buffered I/O
committed = true;
}
unlock();
return committed;
}

Figure 6. Pseudo-code for atomic commit.

To provide the appearance of sequential execution, Grace
not only needs to provide isolation of each thread, but also
must enforce a particular commit order. Grace employs a
simple commit algorithm that provides the effect of a sequential execution.
Graces commit algorithm implements the following policy: a thread is only allowed to commit after all of its logical predecessors have completed. It might appear that such a
commit protocol would be costly to implement, possibly re-

void * spawnThread (threadFunction * fn,

void * arg) {
// End atomic section here.
atomicEnd();
// Allocate shared mem object
// to hold threads return value.
ThreadStatus * t =
new (allocateStatus()) ThreadStatus;
// Use fork instead of thread spawn.
int child = fork();
if (child) {
// Im the parent (caller of spawn).
// Store the tid to allow later sync
// on child thread.
t->tid = child;
// The spawned child is new predecessor.
predecessor = child;
// Start new atomic section
// and return thread info.
atomicBegin();
return (void *) t;
} else {
// Im the child.
// Set thread id.
tid = getpid();
// Execute thread function.
atomicBegin();
t->retval = fn(arg);
atomicEnd();
// Indicate that process has ended
// to alert its successor (parent)
// that it can continue.
setExited();
// Done.
_exit (0);
}
}

Figure 7. Pseudo-code for thread creation. Note that the

actual Grace library wraps thread creation and joining with
a pthreads-compatible API.
quiring global synchronization and complex data structures.
Instead, Grace employs a simple and efficient commit algorithm, which threads the tree of dependencies through all the
executing threads to ensure sequential semantics.
Executing threads form a tree, where the post-order
traversal specifies the correct commit order. Parents must
wait for their last-spawned child, children wait either for
their preceding sibling if it exists, or the parents previous
sibling. Grace threads the tree of dependencies through all
the executing threads to ensure sequential semantics.
The key is that only thread spawns affect commit dependence, and then only affect those of the newly-spawned child
and parent processes. Each new child always appears immediately before its parent in the post-order traversal. Updating the predecessor values is akin to inserting the child pro-

void joinThread (void * v, void ** result) {

ThreadStatus * t = (ThreadStatus *) v;
// Wait for a particular thread
// (if argument non-NULL).
if (v != NULL) {
atomicEnd();
// Wait for thread to terminate.
if (t->tid)
waitExited (t->tid);
// Grab thread result from status.
if (result != NULL) {
*result = t->retval;
// Reclaim memory.
freeStatus(t);
}
atomicBegin();
}
}

its immediate predecessor to commit. Grace then checks to

make sure that its current state is consistent with the committed state. Once both of these conditions are met, the current thread is then guaranteed to commit when it terminates.
Grace then allows the thread to perform the irrevocable I/O
operation, which is now safe because the threads execution
is guaranteed to succeed.

Figure 8. Pseudo-code for thread joining.

We evaluate Graces performance on real computation kernels with a range of benchmarks, listed in Table 2. One
benchmark, matmula recursive matrix-matrix multiply routinecomes from the Cilk distribution. We handtranslated this program to use the pthreads API (essentially replacing Cilk calls like spawn with their counterparts). We performed the same translation for the remaining Cilk benchmarks, but because they use unusually fine-grained threads, none of them scaled when using
pthreads.
The remaining benchmarks are from the Phoenix benchmark suite [34]. These benchmarks represent kernel computations and were designed to be representative of computeintensive tasks from a range of domains, including enterprise
computing, artificial intelligence, and image processing. We
use the pthreads-based variants of these benchmarks with
the largest available inputs.
In addition to describing the benchmarks, Table 2 also
presents detailed benchmark characteristics measured from
their execution with Grace, including the total number of
commits and rollbacks, together with the average number
of pages read and written and average wall-clock time per
atomic region. With the exception of matmul and kmeans,
the benchmarks read and write from relatively few pages in
each atomic region. matmul has a coarse grain size and
large footprint, but has no interference between threads due
to the regular structure of its recursive decomposition. On
the other hand, kmeans has a benign race which forces
Grace to trigger numerous rollbacks (see Section 6.1).

cess into a linked list representing this traversal. Each child

sets its predecessor to the parents predecessor (which happens automatically because of the semantics of fork), and
then the parent sets its predecessor to the childs ID (see Figure 7).
The parent then continues execution until the next commit point (the end of the thread, a new thread spawn, or when
it joins another thread). At this time, if the parent thread has
read any memory from the heap or globals (see Section 3.4),
it then waits on a semaphore that the child thread sets when
it exits (see Figures 7 and 8).
4.1

Transactional I/O

Graces commit protocol not only enforces sequential semantics but also has an additional important benefit. Because
Grace imposes an order on thread commits, there is always
one thread running that is guaranteed to be able to commit its
state: the earliest thread in program order. This property ensures that Grace programs cannot suffer from livelock caused
by a failure of any thread to make progress, a problem with
some transactional memory systems.
This fact allows Grace to overcome an even more important limitation of most proposed transactional memory systems: it enables the execution of I/O operations in a system
with optimistic concurrency. Because some I/O operations
are irrevocable (e.g., network reads after writes), most I/O
operations appear to be fundamentally at odds with speculative execution. The usual approach is to ban I/O from speculative execution, or to arbitrarily pick a winner to obtain
a global lock prior to executing its I/O operations.
In Grace, each thread buffers its I/O operations and commits them at the same time it commits its updates to memory,
as shown in Figure 6. However, if a thread attempts to execute an irrevocable I/O operation, Grace forces it to wait for

5.

Methodology

We perform our evaluation on a quiescent 8-core system

(dual processor with 4 cores), and 8GB of RAM. Each processor is a 4-core 64-bit Intel Xeon running at 2.33 Ghz
with a 4MB L2 cache. We compare Grace to the Linux
pthreads library (NPTL), on Linux 2.6.23 with GNU libc
version 2.5.
5.1

5.1.1

CPU-Intensive Benchmarks

Modifications

All of these programs run correctly with Grace out of the

box, but as we explain below, they required slight tweaking to allow them to scale (with no modifications, none
of the programs scale). These changes were typically short
and local, requiring one or two lines of new code, and required no understanding of the application itself. Several of

Benchmark
histogram
kmeans
linear regression
matmul
pca
string match

Description
Analyzes images RGB components
Iterative clustering of 3-D points
Computes best fit line for set of points
Recursive matrix-multiply
Principal component analysis on matrix
Searches file for encrypted word

Commits

Rollbacks

9
6273
9
11
22
11

0
4887
0
0
0
0

(average per atomic region)

Pages Read Pages Written Runtime (ms)

7.3
404.5
5.6
4100
3.1
5.9

5.9
2.3
4.8
1865
2.2
4.3

1512.3
8.7
1024.0
2359.4
0.204
191.1

Table 2. CPU-intensive multithreaded benchmark suite and detailed characteristics (see Section 5.1).
CPUintensivebenchmarks
pthreads

Grace

8
7
12.97
6
10.80

Speedu
up

these changes could be mechanically applied by a compiler,

though we have not explored this. (We note that the modification of benchmarks to explore new programming models
is standard practice, e.g., in papers exploring software transactional memory or map-reduce.)
Thread-creation hoisting / argument padding: In most
of the applications, the only modification we made was to
the loop that spawned threads. In the Phoenix benchmarks,
this loop body typically initializes each threads arguments
before spawning the thread. False sharing on these updates
causes Grace to serialize all of threads, precluding scalability. We resolved this either by hoisting the initialization
(initializing thread arguments first in a separate loop and
then spawning the threads), or, where possible, by padding
the thread argument data structures to 4K. In one case, for
the kmeans benchmark, the benchmark erroneously reuses
the same thread arguments for each thread, which not only
causes Grace to serialize the program but also is a race condition. We fixed the code by creating a new heap-allocated
structure to hold the arguments for each thread.
Page-size base case: We made a one-line change to the
matmul benchmark, where we increased the base matrix
size of the recursion to a multiple of the size of a page to prevent false sharing. Interestingly, this modification was beneficial not only for Grace but also for the pthread version. It
not only reduces false sharing across the threads but also improves the baseline performance of the benchmark by around
8% by improving its cache utilization.
Changed concurrency structure: Our most substantial
change (16 lines of code) was to pca, where we changed
the way that the program manages concurrency. The original benchmark divided work dynamically across a number
of threads, with each thread updating a global variable to indicate which row of a matrix to process next: with Grace,
the first thread performed all of the computations. To enable
pca to scale, we statically partitioned the work by providing each thread with a range of rows. This modification had
little impact on the pthreads version but dramatically improved the scalability with Grace.
Summary: The vast majority of the code changes were
local, purely mechanical and required minimal programmer
intervention, primarily in the thread creation loop. In almost
every case, the modifications required no knowledge of the

4
3
2
1
0
histogram

kmeans

linear_regression

matmul

pca

string_match

Benchmarks

Figure 9. Performance of multithreaded benchmarks running with pthreads and Grace on an 8 core system (higher
is better). Grace generally performs nearly as well as the
pthreads version while ensuring the absence of concurrency errors.
underlying application. The reordering or modification involved a small number of lines of code (116).

6.

Evaluation

Our evaluation answers the following questions:

1. How well does Grace perform on real applications?
2. What kind of applications work best with Grace?
3. How effective is Grace against a range of concurrency
errors?
6.1

Real Applications

Figure 9 shows the result of running our benchmark suite of

applications, graphed as their speedup over a serial execution. The Grace-based versions achieve comparable performance while at the same time guaranteeing the absence of
concurrency errors. The average speedup for Grace is 6.2X,
while the average speedup for pthreads is 7.13X.
There are two notable outliers. The first one is pca,
which exhibits superlinear speedups both for Grace and
pthreads. The superlinear speedup is due to improved
cache locality caused by the division of the computation
into smaller chunks across multiple threads.

(b) Impact of grain size (normalized to pthread)

(a) Impact of grain size (speedup)

10
Speedup against sequential execution

Grace
pthread

Grace
pthread

1.4
Normalized Execution Time

1.2
1
0.8
0.6
0.4
0.2
0

0
1

16
32
64
Thread length (ms)

128

256

512

1024

8
16
32
64
128
Thread Execution Length (ms)

256

512

1024

Figure 10. Impact of thread running time on performance: (a) speedup over a sequential version (higher is better), (b)
normalized execution time with respect to pthreads (lower is better).
(b) Impact of footprint (normalized to pthread)

(a) Impact of footprint (speedup)

4

8
7

Thread Size: 50ms

6
5
Thread Size: 10ms

4
3

Grace (10ms)
pthread (10ms)
Grace (50ms)
pthread: (50ms)
Grace: (200ms)
pthread: (200ms)

2
1

16
64
256
Number of pages dirtied (in logscale)

3
2.5
2
1.5
1
0.5
0

0
1

Grace: thread size (10ms)

Grace: thread size (50ms)
Grace: thread size (200ms)
pthread

3.5

Thread Size: 200ms

Normalized Execution Time

Speedup against sequential execution

1024

16
64
256
Number of pages dirtied (in logscale)

1024

Figure 11. Impact of thread running time on performance: (a) speedup over a sequential version (higher is better), (b)
normalized execution time with respect to pthreads (lower is better).
While the kmeans benchmark achieves a modest speedup
with pthreads (3.65X), it exhibits no speedup with Grace
(1.02X), which serializes execution. This benchmark iteratively clusters points in 3D space. Until it makes no further
modifications, kmeans spawns threads to find clusters (setting a cluster id for each point), and then spawns threads to
compute and store mean values in a shared array. It would be
straightforward to eliminate all rollbacks for the first threads
by simply rounding up the number of points assigned to
each thread, allowing each thread to work on independent
regions of memory. However, kmeans does not protect accesses or updates to the mean value array and instead uses
benign races as a performance optimization. Grace has no
way of knowing that these races are benign and serializes its
execution to prevent the races.
6.2

Application Characteristics

While the preceding evaluation shows that Grace performs

well on a range of benchmarks, we also developed a microbenchmark to explore a broader space of applications. In
particular, our microbenchmark allows us to vary the follow-

ing parameters: grain size, the running time of each thread;

footprint, the number of pages updated by a thread; and conflict rate, the likelihood of conflicting updates by a thread.
These parameters isolate Graces overheads. First, the
shorter a threads execution (the smaller its grain), the more
the increased cost of thread spawns in Grace (actually process creation) should dominate. Second, increasing the number of pages accessed by a thread (its footprint) stresses the
cost of Graces page protection and signal handling. Third,
increasing the number of conflicting updates forces Grace to
rollback and re-execute code more often, degrading performance.
Grain size: We first evaluate the impact of the length
of thread execution on Graces performance. We execute
a range of tests, where each thread runs for some fixed
number of milliseconds performing arithmetic operations
in a tight loop. Notice that this benchmark only exercises
the CPU and the cost of thread creation and destruction,
because it does not reference heap pages or global data. Each
experiment is configured to run for a fixed amount of time:
nTh len nIter = 16 seconds, where nTh is the number of

Impact of Conflict Rate

10
Grace
Pthread
8

Speedup

threads (16), len is the thread running time, and nIter is the
number iterations.
Figure 10 shows the effect of thread running time on
performance. Because we expected the higher cost of thread
spawns to degrade Graces performance relative to pthreads,
we were surprised to view the opposite effect. We discovered
that the operating systems scheduling policy plays an important role in this set of experiments.
When the size of each thread is extremely small, neither
Grace nor pthreads make effective use of available CPUs.
In both cases, the processes/threads finish so quickly that the
load balancer is not triggered and so does not run them on
different CPUs. As the thread running time becomes larger,
Grace tends to make better of CPU resources, sometimes up
to 20% faster. We believe this is because the Linux CPU
scheduler attempts to put threads from the same process on
one CPU to exploit cache locality, which limits its ability to
use more CPUs, but is more liberal in its placement of processes across CPUs. However, once thread running time becomes large enough (over 50ms) for the load balancer to take
effect, both Grace and pthreads scale well. Figure 10(b)
shows that Grace has competitive performance compared to
pthreads, and the overhead of process creation is never
larger than 2%.
Footprint: In order to evaluate the impact of per-thread
footprint, we extend the previous benchmark so that each
thread also writes a value onto a number of private pages,
which only exercises Graces page protection mechanism
without triggering rollbacks. We conduct an extensive set of
tests, ranging thread footprint from 1 pages to 1024 pages
(4MB). This experiment is the worst case scenario for Grace,
since each write triggers two page faults.
Figure 11 summarizes the effect of thread footprint over
three representative thread running time settings: small
(10ms), medium (50ms) and large (200ms). When the thread
footprint is not too large ( 64 pages), Grace has comparable performance to pthreads, with no more than a 5%
slowdown. As the thread footprint continues to grow, the
performance of Grace starts to degrade due the overhead
of page protection faults. However, even when each thread
dirties one megabyte of memory (256 pages), Graces performance is within an acceptable range for the medium and
large thread runtime settings. The overhead of page protection faults only becomes prohibitively large when the thread
footprint is large relative to the running time, which is unlikely to be representative of compute-intensive threads.
Conflict rate: We next measure the impact of conflicting
updates on Graces performance by having each thread in the
microbenchmark update a global variable with a given probability, which the result that any other thread reading or writing that variable will need to rollback and re-execute. Grace
makes progress even with a 100% likelihood of conflicts because its sequential semantics provide a progress guarantee:
the first thread in commit order is guaranteed to succeed

0
0

20

40
60
Conflict Rate (%)

80

100

Figure 12. Impact of conflict rate (the likelihood of conflicting updates, which force rollbacks), versus a pthreads
baseline that never rolls back (higher is better).

without rolling back Figure 12 shows the resulting impact

on speedup (where each thread runs for 50 milliseconds).
When the conflict rate is low, Graces performance remains close to that of pthreads. Higher conflict rates degrade Graces performance, though to a diminishing extent:
a 5% conflict rate leads to a 6-way speedup, while a 100%
conflict rate matches the performance of a serial execution.
In this benchmark, one processor is always performing useful work, so performance matches the serial baseline. In a
program with many more threads than processors, however,
a 100% conflict rate under Grace would result in a slowdown.
Summary: This microbenchmark demonstrates that the
use of processes versus threads in Grace has little impact on
performance for threads that run as little as 10ms, adding
no more than 2% overhead and actually providing slightly
better scalability than pthreads in some cases. Memory
protection overhead is minimal when the number of pages
dirtied is not excessively large compared to the grain size
(e.g., up to 2MB for 50ms threads). Rollbacks triggered
by conflicting memory updates have the largest impact on
performance. While Grace can provide scalability for high
conflict rates, the conflict rate should be kept relatively low
to ensure reasonable performance relative to pthreads.
6.3

Concurrency Errors

We illustrate Graces ability to eliminate most concurrency

bugs by compiling a bug suite primarily drawn from actual
bugs described in previous work on error detection and listed
in Table 3 [25, 26, 27]. Because concurrency errors are by
their nature non-deterministic and occur only for particular
thread interleavings, we inserted delays (via the usleep
function call) at key points in the code. These delays dramatically increase the likelihood of encountering these errors, allowing us to compare the effect of using Grace and
pthreads.

Bug type
deadlock
race condition
atomicity violation
order violations

Benchmark description
Cyclic lock acquisition
Race condition example, Lucia et al. [27]
Atomicity violation from MySQL [26]
Order violation from Mozilla 0.8 [25]

Table 3. Error benchmark suite.

// Deadlock.
thread1 () {
lock (A);
// usleep();
lock (B);
// ...do something
unlock (B);
unlock (A);
}
thread2 () {
lock (B);
// usleep();
lock (A);
// ...do something
unlock (A);
unlock (B);
}

Figure 13. Deadlock example. This code has a cyclic

lock acquisition pattern that triggers a deadlock under
pthreads while running to completion with Grace.

6.3.1

Deadlocks

Figure 13 illustrates a deadlock error caused by cyclic lock

acquisition. This example spawns two threads that each attempt to acquire two locks A and B, but in different orders:
thread 1 acquires lock A then lock B, while thread 2 acquires lock B then lock A. When using pthreads, these
threads deadlock if both of them manage to acquire their first
locks, because each of the threads is waiting to acquire a lock
held by the other thread. Inserting usleep after these locks
makes this program deadlock reliably under pthreads.
However, because Graces atomicity and commit protocol
lets it treat locks as no-ops, this program never deadlocks
with Grace.
6.3.2

Race conditions

We next adapt an example from Lucia et al. [27], removing

the lock in the original example to trigger a race. Figure 14
shows two threads both executing increment, which increments a shared variable counter. However, because access to counter is unprotected, both threads could read the
same value and so can lose an update. Running this example
under pthreads with an injected delay exhibits this race,
printing 0,0,1,1. By contrast, Grace prevents the race by

// Race condition.
int counter = 0;
increment() {
print (counter);
int temp = counter;
temp++;
// usleep();
counter = temp;
print (counter);
}
thread1() { increment(); }
thread2() { increment(); }
}

Figure 14. Race condition example: the race is on the variable counter, where the first update can be lost. Under
Grace, both increments always succeed.
// Atomicity violation.
// thread1
S1: if (thd->proc_info) {
// usleep();
S2:
fputs (thd->proc_info,..)
}
// thread2
S3: thd->proc_info = NULL;

Figure 15. An atomicity violation from MySQL [26]. A

faulty interleaving can cause this code to trigger a segmentation fault due to a NULL dereference, but by enforcing
atomicity, Grace prevents this error.
executing each thread deterministically, and invariably outputs the sequence 0,1,1,2.
6.3.3

Atomicity Violations

To verify Graces ability to cope with atomicity violations,

we adapted an atomicity violation bug taken from MySQLs
InnoDB module, described by Lu et al. [26]. In this example,
shown in Figure 15, the programmer has failed to properly
protect access to the global variable thd. If the scheduler
executes the statement labeled S3 in thread 2 immediately
after thread 1 executes S1, the program will dereference
NULL and fail.
Inserting a delay between S1 and S2 causes every execution of this code with pthreads to segfault because of
a NULL dereference. With Grace, threads appear to execute
atomically, so the program always performs correctly.
6.3.4

Order violations

Finally, we consider order violations, which were recently

identified as a common class of concurrency errors by Lu et

// Order violation.
char * proc_info;

// Order violation.
int foo;

thread1() {
// ...
// usleep();
proc_info = malloc(256);
}

thread1() {
foo = 0;
}

thread2() {
// ...
strcpy(proc_info,"abc");
}
main() {
spawn thread1();
spawn thread2();
}

Figure 16. An order violation. If thread 2 executes before

thread 1, it writes into unallocated memory. Grace ensures
that thread 2 always executes after thread 1, avoiding this
error.

al. [26]. An order violation occurs when the program runs

correctly under one ordering of thread executions, but incorrectly under a different schedule. Notice that order violations
are orthogonal to atomicity violations: an order violation can
occur even when the threads are entirely atomic.
Figure 16 presents a case where the programmers intended order is not guaranteed to be obeyed by the scheduler.
Here, if thread 2 manages to write into proc info before it
has been allocated by thread 1, it will cause a segfault. However, because the scheduler is unlikely to be able to schedule thread 2 before thread 1 has executed the allocation call,
this code will generally work correctly. Nonetheless, it will
occasionally fail, and injecting usleep() forces it to fail
reliably. With Grace, this microbenchmark always runs correctly, because Grace ensures that the spawned threads exhibit sequential semantics. Thus, thread 2 can commit only
after thread 1 completes, preventing the order violation.
Interestingly, while Grace prescribes the order of program
execution, Figure 17 shows that the expected order might not
be the order that Grace enforces. In this example, modeled
after an order violation bug from Mozilla, the pthreads
version is almost certain to execute statement S2 immediately after S1; that is, well before the scheduler is able to
run thread1. The final value of foo (once thread1 executes) will therefore almost always be 0.
However, in the rare event that a context switch occurs
immediately after S1, the thread may get a chance to run
first, leaving the value of foo at 1 and causing the assertion
to fail. Such a bug would be unlikely to be revealed during
testing and could lead to failures in the field that would be
exceedingly difficult to locate.

main() {
S1: spawn thread1();
// usleep();
S2: foo = 1;
// ...
assert (foo == 0);
}

Figure 17. An order violation. Here, the intended effect

violates sequential semantics, so the error is not fixed but
occurs reliably.
However, with Grace, the final value of foo will always
be 1, because that result corresponds to the result of a sequential execution of thread1. While this result might
not have been the one that the programmer expected, using
Grace would have made the error both obvious and repeatable, and thus easier to fix.

7.

Related Work

The literature relating to concurrent programming is vast.

We briefly describe the most closely-related work here.
7.1

Transactional memory

The area of transactional memory, first proposed by Herlihy

and Moss for hardware [17] and for software by Shavit and
Touitou [36], is now a highly active area of research. Larus
and Rajwars book provides an overview of recent work
in the area [23]. We limit our discussion here to the most
closely related software approaches that run on commodity
hardware.
Transactional memory eliminates deadlocks but does not
address other concurrency errors like races and atomicity,
leaving the burden on the programmer to get the atomic
sections right. Worse, software-based transactional memory systems (STM) typically interact poorly with irrevocable operations like I/O and generally degrade performance
when compared to their lock-based counterparts, especially
those that provide strong atomicity [6]. STMs based on weak
atomicity can provide reasonable performance but expose
programmers to a range of new and subtle errors [37].
Fraser and Harriss transaction-based atomic blocks [15]
are a programming construct that has been the model for
many subsequent language proposals. However, the semantics of these language proposals are surprisingly complex.
For example, Shpeisman et al. [37] show that proposed
weak transactions can give rise to unanticipated and unpredictable effects in programs that would not have arised

when using lock-based synchronization. With Grace, program semantics are straightforward and unsurprising.
Welc et al. introduce support for irrevocable transactions
in the McRT-STM system for Java [40]. Like Grace, their
system supports one active irrevocable transaction at a time.
McRT-STM relies on a lock mechanism combined with
compiler-introduced read and write barriers, while Graces
support for I/O falls out for free from its commit protocol.
The McRT system for C++ also includes a malloc implementation called McRT-malloc, which resembles Hoard [3]
but is extended to support transactions [19]. Ni et al. present
the design and implementation of a transactional extension
to C++ that enable transactional use of the system memory
allocator by wrapping all memory management functions
and providing custom commit and undo actions [31]. These
approaches differ substantially from Graces memory allocator, which employs a far simpler design that leverages the
fact that in Grace, all code, including malloc and free,
execute transactionally. Grace also takes several additional
steps that reduce the risk of false sharing.
7.2

Concurrent programming models

We restrict our discussion of programming models here to

imperative rather than functional programming languages.
Cilk [5] is a multithreaded extension of the C programming
language. Like Grace, Cilk uses a fork-join model of parallelism and focuses on the use of multiple threads for CPU
intensive workloads, rather than server applications. Unlike
Grace, which works with C or C++ binaries, Cilk is currently
restricted to C. Cilk also relies on programmers to avoid
race conditions and other concurrency errors; while there
has been work on dynamic tools to locate these errors [8],
Grace automatically prevents them. A proposed variant of
Cilk called Transactions Everywhere adds transactions to
Cilk by having the compiler insert cutpoints (transaction end
and begin) at various points in the code, including at the
end of loop iterations. While this approach reduces exposure to concurrency errors, it does not prevent them, and data
race detection in this model has been shown to be an NPcomplete problem [18]. Concurrency errors remain common
even in fork-join programs: Feng and Leiserson report that
their Nondeterminator race detector for Cilk found races in
several Cilk programs written by experts, as well as in half
the submitted implementations of Strassens matrix-matrix
multiply in a class at MIT [13].
Intels Threading Building Blocks (TBB) is a C++ library
that provides lightweight threads (tasks) executing on a
Cilk-like runtime system [35]. TBB comprises a non-POSIX
compatible API, primarily building on a fork-join programming model with concurrent containers and high-level loop
constructs like parallel do that abstract away details
like task creation and barrier synchronization (although TBB
also includes support for pipeline-based parallelism, which
Grace does not). TBB relies on the programmer to avoid concurrency errors that Grace prevents.

Automatic mutual exclusion, or AME, is a recentlyproposed programming model developed at Microsoft Research Cambridge. It is a language extension to C# that assumes that all shared state is private unless otherwise indicated [20]. These guarantees are weaker than Graces, in that
AME programmers can still generate code with concurrency
errors. AME has a richer concurrent programming model
than Grace that makes it more flexible, but its substantially
more complex semantics preclude a sequential interpretation [1]. By contrast, Graces semantics are straightforward
and thus likely easier for programmers to understand.
von Praun et al. present Implicit Parallelism with Ordered
Transactions (IPOT), that describes a programming model,
like Grace, that supports speculative concurrency and enforces determinism [38]. However, unlike Grace, IPOT requires a completely new programming language, with a wide
range of constructs including variable type annotations and
constructs to support speculative and explicit parallelism. In
addition, IPOT would require special hardware and compiler
support, while Grace operates on existing C/C++ programs
that use standard thread constructs.
Welc et al. present a future-based model for Java programming that, like Grace, is safe [39]. A future denotes
an expression that may be evaluated in parallel with the rest
of the program; when the program uses the expressions
value, it waits for the future to complete execution before
continuing. As with Graces threads, safe futures ensure that
the concurrent execution of futures provides the same effect
as evaluating the expressions sequentially. However, the safe
future system assumes that writes are rare in futures (by contrast with threads), and uses an object-based versioning system optimized for this case. It also requires compiler support
and currently requires integration with a garbage-collected
environment, making it generally unsuitable for use with
C/C++.
Graces use of virtual memory primitives to support speculation is a superset of the approach used by behaviororiented parallelism (BOP) [12]. BOP allows programmers
to specify possibly parallelizable regions of code in sequential programs, and uses a combination of compiler analysis
and the strong isolation properties of processes to ensure that
speculative execution never prevents a correct execution.
While BOP seeks to increase the performance of sequential code by enabling safe, speculative parallelism, Grace
provides sequential semantics for concurrently-executing,
fork-join based multithreaded programs.
7.3

Deterministic thread execution

A number of runtime systems have recently appeared that

are designed to provide a measure of deterministic execution of multithreaded programs. Isolator uses a combination
of programmer annotation, custom memory allocation, and
virtual memory primitives to ensure that programs follow
a locking discipline [33]. Isolator works on existing lockbased codes, but does not address issues like atomicity or

deadlock. Kendo also works on stock hardware and provides

deterministic execution, but only of the order of lock acquisitions [32]. It also requires data-race free programs. DMP
uses hardware support to provide a total ordering on multithreaded execution, which aims to ensure that programs reliably exhibit the same errors, rather than attempting to eliminate concurrency errors altogether [10].
In concurrent work, Bocchino et al. present Deterministic
Parallel Java (DPJ), a dialect of Java that adds two parallel
constructs (cobegin and foreach) [21]. A programmer
using DPJ provides region annotations to describe accesses
to disjoint regions of the heap. DPJs type and effect system
then verifies the soundness of these annotations at compiletime, allowing it to execute non-interfering code in parallel
with the guarantee that the parallel code executes with the
same semantics as a sequential execution (although it relies on the correctness of commutativity annotations). Unlike Grace, DPJ does not rely on runtime support, but requires programmer-supplied annotations and cannot provide
correctness guarantees for ordinary multithreaded code outside the parallel constructs.
7.4

Other uses of virtual memory

A number of distributed shared memory (DSM) systems

of the early 90s also employed virtual memory primitives
to detect reads and writes and implement weaker consistency models designed to improve DSM performance, including Munin [7] and TreadMarks [22]. While both Grace
and these DSM systems rely on these mechanisms to trap
reads and writes, the similarities end there. Grace executes
multithreaded shared memory programs on shared memory
systems, rather than creating the illusion of shared memory
on a distributed system, where the overheads of memory protection and page fault handling are negligible compared to
the costs of network transmission of shared data.

8.

Future Work

In this section, we outline several directions for future work

for Grace, including extending its range of applicability and
further improving performance.
We intend to extend Grace to support other models of
concurrency beyond fork/join parallelism. One potential
class of applications is request/response servers, where a single controller thread spawns many mostly-indepedent child
threads. For these programs, Grace could guarantee isolation for child threads while maintaining scalability. This approach would require modifying Graces semantics to allow
the controller thread to spawn new children without committing in order to allow it to handle the side-effects of socket
communication without serializing spawns of child threads.
While conflicts cause rollbacks, they also provide potentially useful information that can be fed back into the runtime system. We are building enhancements to Grace that
will both report memory areas that are the source of frequent

conflicts and act on this information. This information can

guide programmers as they tune their programs for higher
performance. More importantly, we are currently developing a tool that will allow this data to be used by Grace to
automatically prevent conflicts (without programmer intervention) by padding or segregating conflicting heap objects
from different call sites.
While we have shown that process invocation is surprisingly efficient, we would like to further reduce the cost of
threads. While we do not evaluate it here, we recently developed a technique that greatly lowers the cost of thread
invocation by taking advantage of the following key insight.
Once a divide-and-conquer application has spawned a large
enough number of threads to take advantage of available
processors, it is possible to practically eliminate the cost of
thread invocation at deeper nesting levels by directly executing thread functions instead of spawning new processes.
While this approach has no impact on our benchmark suite,
it dramatically decreases the cost of thread spawns, running
at under 2X the cost of Cilks lightweight threads.
Another possible use of rollback information would be
for scheduling: the runtime system could partition threads
into conflicting sets, and then only schedule the first thread
(in serial order) from each of these sets. This algorithm
would maximize the utilization of available parallelism by
preventing repeated rollbacks.
We are also investigating the use of compiler optimizations to automatically transform code to increase scalability. For example, Graces sequential semantics could enable cross-thread optimizations, such as hoisting conflicting
memory operations out of multiple threads.

9.

Conclusion

This paper presents Grace, a runtime system for fork-join

based C/C++ programs that, by replacing the standard
threads library with a system that ensures deterministic
execution, eliminates a broad class of concurrency errors,
including deadlocks, race conditions, atomicity violations,
and order violations. With modest source code modifications (116 lines of code in our benchmark suite), Grace
generally achieves good speed and scalability on multicore
systems while providing safety guarantees. The fact that
Grace makes multithreaded program executions deterministic and repeatable also has the potential to greatly simplify
testing and debugging of concurrent programs, even where
deploying Grace might not be feasible.

10.

Acknowledgements

The authors would like to thank Ben Zorn for his feedback
during the development of the ideas that led to Grace, to
Luis Ceze for graciously providing benchmarks, and to Cliff
Click, Dave Dice, Sam Guyer, and Doug Lea for their invaluable comments on earlier drafts of this paper. We also
thank Divya Krishnan for her assistance. This material is

based upon work supported by Intel, Microsoft Research,

and the National Science Foundation under CAREER Award
CNS-0347339 and CNS-0615211. Any opinions, findings,
and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect
the views of the National Science Foundation.

References

[11] D. Dice, O. Shalev, and N. Shavit. Transactional locking ii.

In S. Dolev, editor, DISC, volume 4167 of Lecture Notes in
Computer Science, pages 194208. Springer, 2006.
[12] C. Ding, X. Shen, K. Kelsey, C. Tice, R. Huang, and C. Zhang.
Software behavior oriented parallelization. In PLDI 07:
Proceedings of the 2007 ACM SIGPLAN conference on
Programming language design and implementation, pages
223234, New York, NY, USA, 2007. ACM.

[1] M. Abadi, A. Birrell, T. Harris, and M. Isard. Semantics of

transactional memory and automatic mutual exclusion. In
POPL 08: Proceedings of the 35th annual ACM SIGPLANSIGACT symposium on Principles of programming languages,
pages 6374, New York, NY, USA, 2008. ACM.

[13] M. Feng and C. E. Leiserson. Efficient detection of

determinacy races in cilk programs. In SPAA 97: Proceedings
of the ninth annual ACM symposium on Parallel algorithms
and architectures, pages 111, New York, NY, USA, 1997.
ACM.

[2] D. F. Bacon and S. C. Goldstein. Hardware-assisted replay

of multiprocessor programs. In PADD 91: Proceedings
of the 1991 ACM/ONR workshop on Parallel and distributed
debugging, pages 194206, New York, NY, USA, 1991. ACM.

[14] C. Flanagan and S. Qadeer. A type and effect system for

atomicity. In PLDI 03: Proceedings of the ACM SIGPLAN
2003 conference on Programming language design and
implementation, pages 338349, New York, NY, USA, 2003.
ACM.

[3] E. D. Berger, K. S. McKinley, R. D. Blumofe, and P. R. Wilson. Hoard: A scalable memory allocator for multithreaded
applications. In Proceedings of the International Conference
on Architectural Support for Programming Languages and
Operating Systems (ASPLOS-IX), pages 117128, New York,
NY, USA, Nov. 2000. ACM.
[4] E. D. Berger, B. G. Zorn, and K. S. McKinley. Composing
high-performance memory allocators. In Proceedings of the
2001 ACM SIGPLAN Conference on Programming Language
Design and Implementation (PLDI 2001), pages 114124,
New York, NY, USA, June 2001. ACM.
[5] R. D. Blumofe, C. F. Joerg, B. C. Kuszmaul, C. E. Leiserson,
K. H. Randall, and Y. Zhou. Cilk: an efficient multithreaded
runtime system. J. Parallel Distrib. Comput., 37(1):5569,
1996.
[6] C. Blundell, E. C. Lewis, and M. M. K. Martin. Deconstructing transactions: The subtleties of atomicity. In WDDD 05:
4th Workshop on Duplicating, Deconstructing, and Debunking, June 2005.
[7] J. B. Carter, J. K. Bennett, and W. Zwaenepoel. Implementation and performance of munin. In SOSP 91: Proceedings of
the Thirteenth ACM Symposium on Operating Systems Principles, pages 152164, New York, NY, USA, 1991. ACM.
[8] G.-I. Cheng, M. Feng, C. E. Leiserson, K. H. Randall, and
A. F. Stark. Detecting data races in cilk programs that use
locks. In SPAA 98: Proceedings of the tenth annual ACM
symposium on Parallel algorithms and architectures, pages
298309, New York, NY, USA, 1998. ACM.
[9] J. Dean and S. Ghemawat. MapReduce: simplified data
processing on large clusters. In OSDI04: Proceedings of the
6th conference on Symposium on Opearting Systems Design
& Implementation, pages 1010, Berkeley, CA, USA, 2004.
USENIX Association.
[10] J. Devietti, B. Lucia, L. Ceze, and M. Oskin. DMP:
deterministic shared memory multiprocessing. In ASPLOS
09: Proceedings of the 14th International Conference on
Architectural Support for Programming Languages and
Operating Systems, pages 8596, New York, NY, USA, 2009.
ACM.

[15] T. Harris and K. Fraser. Language support for lightweight

transactions. In OOPSLA 03: Proceedings of the 18th annual
ACM SIGPLAN conference on Object-oriented programing,
systems, languages, and applications, pages 388402, New
York, NY, USA, 2003. ACM.
[16] J. W. Havender. Avoiding deadlock in multitasking systems.
IBM Systems Journal, 7(2):7484, 1968.
[17] M. Herlihy and J. E. B. Moss. Transactional memory:
architectural support for lock-free data structures. In ISCA
93: Proceedings of the 20th annual international symposium
on Computer architecture, pages 289300, New York, NY,
USA, 1993. ACM.
[18] K. Huang. Data-race detection in transactions-everywhere
parallel programming. Masters thesis, Department of
Electrical Engineering and Computer Science, Massachusetts
Institute of Technology, June 2003.
[19] R. L. Hudson, B. Saha, A.-R. Adl-Tabatabai, and B. C.
Hertzberg. Mcrt-malloc: a scalable transactional memory
allocator. In ISMM 06: Proceedings of the 5th International
Symposium on Memory Management, pages 7483, New
York, NY, USA, 2006. ACM.
[20] M. Isard and A. Birrell. Automatic mutual exclusion. In
HotOS XI: 11th Workshop on Hot Topics in Operating
Systems, Berkeley, CA, May 2007.
[21] R. L. B. Jr., V. S. Adve, D. Dig, S. Adve, S. Heumann,
R. Komuravelli, J. Overbey, P. Simmons, H. Sung, and
M. Vakilian. A type and effect system for deterministic
parallel Java. In OOPSLA 09: Proceedings of the 24th
ACM SIGPLAN Conference on Object-oriented Programming
Systems, Languages, and Applications, New York, NY, USA,
2009. ACM.
[22] P. Keleher, A. L. Cox, S. Dwarkadas, and W. Zwaenepoel.
Treadmarks: Distributed shared memory on standard workstations and operating systems. In WTEC94: Proceedings of
the USENIX Winter 1994 Technical Conference, pages 1010,
Berkeley, CA, USA, 1994. USENIX Association.
[23] J. R. Larus and R. Rajwar. Transactional Memory. Morgan &
Claypool, 2006.

[24] D. Lea. A Java fork/join framework. In JAVA 00: Proceedings

of the ACM 2000 conference on Java Grande, pages 3643,
New York, NY, USA, 2000. ACM.
[25] S. Lu, S. Park, C. Hu, X. Ma, W. Jiang, Z. Li, R. A. Popa, and
Y. Zhou. MUVI: automatically inferring multi-variable access
correlations and detecting related semantic and concurrency
bugs. In SOSP 07: Proceedings of the Twenty-First ACM
SIGOPS Symposium on Operating Systems Principles, pages
103116, New York, NY, USA, 2007. ACM.
[26] S. Lu, S. Park, E. Seo, and Y. Zhou. Learning from
mistakes: a comprehensive study on real world concurrency
bug characteristics. In ASPLOS XIII: Proceedings of the
13th international conference on Architectural support for
programming languages and operating systems, pages 329
339, New York, NY, USA, 2008. ACM.
[27] B. Lucia, J. Devietti, K. Strauss, and L. Ceze. Atom-Aid:
Detecting and surviving atomicity violations. In ISCA 08:
Proceedings of the 35th Annual International Symposium on
Computer Architecture, New York, NY, USA, June 2008.
ACM Press.
[28] C.-K. Luk, R. Cohn, R. Muth, H. Patil, A. Klauser, G. Lowney,
S. Wallace, V. J. Reddi, and K. Hazelwood. Pin: building
customized program analysis tools with dynamic instrumentation. In PLDI 05: Proceedings of the 2005 ACM SIGPLAN
conference on Programming language design and implementation, pages 190200, New York, NY, USA, 2005. ACM.
[29] C. E. McDowell and D. P. Helmbold. Debugging concurrent
programs. ACM Comput. Surv., 21(4):593622, 1989.
[30] R. H. B. Netzer and B. P. Miller. What are race conditions?:
Some issues and formalizations. ACM Lett. Program. Lang.
Syst., 1(1):7488, 1992.
[31] Y. Ni, A. Welc, A.-R. Adl-Tabatabai, M. Bach, S. Berkowits, J. Cownie, R. Geva, S. Kozhukow, R. Narayanaswamy,
J. Olivier, S. Preis, B. Saha, A. Tal, and X. Tian. Design
and implementation of transactional constructs for C/C++.
In OOPSLA 08: Proceedings of the 23rd ACM SIGPLAN
Conference on Object-oriented Programming Systems, Languages, and Applications, pages 195212, New York, NY,
USA, 2008. ACM.
[32] M. Olszewski, J. Ansel, and S. Amarasinghe. Kendo:
efficient deterministic multithreading in software. In ASPLOS
09: Proceedings of the 14th International Conference on
Architectural Support for Programming Languages and
Operating Systems, pages 97108, New York, NY, USA,
2009. ACM.
[33] S. Rajamani, G. Ramalingam, V. P. Ranganath, and
K. Vaswani. ISOLATOR: dynamically ensuring isolation
in comcurrent programs. In ASPLOS 09: Proceeding of
the 14th International Conference on Architectural Support
for Programming Languages and Operating Systems, pages
181192, New York, NY, USA, 2009. ACM.

[34] C. Ranger, R. Raghuraman, A. Penmetsa, G. Bradski, and

C. Kozyrakis. Evaluating MapReduce for multi-core and
multiprocessor systems. In Proceedings of the 13th Intl.
Symposium on High-Performance Computer Architecture
(HPCA), feb 2007.
[35] J. Reinders. Intel Threading Building Blocks: Outfitting C++
for Multi-core Processor Parallelism. OReilly Media, Inc.,
2007.
[36] N. Shavit and D. Touitou. Software transactional memory.
In PODC 95: Proceedings of the fourteenth annual ACM
symposium on Principles of distributed computing, pages
204213, New York, NY, USA, 1995. ACM.
[37] T. Shpeisman, V. Menon, A.-R. Adl-Tabatabai, S. Balensiefer,
D. Grossman, R. L. Hudson, K. F. Moore, and B. Saha.
Enforcing isolation and ordering in STM. In PLDI 07:
Proceedings of the 2007 ACM SIGPLAN conference on
Programming language design and implementation, pages
7888, New York, NY, USA, 2007. ACM.
[38] C. von Praun, L. Ceze, and C. Cascaval. Implicit parallelism
with ordered transactions. In PPoPP 07: Proceedings of the
12th ACM SIGPLAN Symposium on Principles and Practice
of Parallel Programming, pages 7989, New York, NY, USA,
2007. ACM.
[39] A. Welc, S. Jagannathan, and A. Hosking. Safe futures for
Java. In OOPSLA 05: Proceedings of the 20th annual ACM
SIGPLAN Conference on Object oriented Programming,
Systems, Languages, and applications, pages 439453, New
York, NY, USA, 2005. ACM.
[40] A. Welc, B. Saha, and A.-R. Adl-Tabatabai. Irrevocable
transactions and their applications. In SPAA 08: Proceedings
of the Twentieth Annual Symposium on Parallelism in
Algorithms and Architectures, pages 285296, New York,
NY, USA, 2008. ACM.