[Image 1]

Introduction

Hey it's a me again @drifter1!

Today we continue with the Parallel Programming series about the OpenMP API. I highly suggest you to go read the previous articles of the series, that you can find by the end of this one. Today we will get into how parallel threads can be synchronized using Locks and Barriers!

So, without further ado, let's get straight into it!

GitHub Repository

https://github.com/drifter1/openmptutorials

Requirements - Prerequisites

Basic understanding of the Programming Language C, or even C++
Familiarity with Parallel Computing/Programming in general
Compiler

Linux users: GCC (GNU Compiler Collection) installed
Windows users: MinGW32/64 - To avoid unnessecary problems I suggest using a Linux VM, Cygwin or even a WSL Environment on Windows 10
MacOS users: Install GCC using brew or use the Clang compiler
For more Compilers & Tools check out: https://www.openmp.org//resources/openmp-compilers-tools/

Previous Articles of the Series

Recap

A parallel region is defined using a parallel region construct:

#pragma omp parallel [clause ...]
{
    /* This code runs in parallel */
}

Parallel For Loops are defined by using the following syntax:

int i;
#pragma omp parallel for private(i) [clause ...]
for(i = 0; i < N; i++){
    ...
}

Non-iterative work-sharing is defined using a sections construct with nested section directives:

#pragma omp sections [clause ...]
{
    /* run in parallel by all threads of the team */

    #pragma omp section
    {
        /* run once by one thread */
    }

    #pragma omp section
    {
        /* run once by one thread */
    } 
    ...
}

When only one section inside a parallel region has to be executed by one thread then we can use the single construct:

#pragma omp single

Critical Sections are specified using the following construct:

#pragma omp critical [[(name)] [ hint (hint-expr)]] [clause ...]
{
    /* critical section code */
}

Atomic operations are defined as:

#pragma omp atomic [read | write | update | capture] [hint(hint-expr)] [clause ...]
    /* statement expression*/

When more-advanced work-sharing is needed, Task constructs are the way to go:

#pragma omp task [clause ...]
{
    /* code block */
}

Locks In OpenMP

Previously in the series, we learned how to define critical sections and atomic operations, which make specific parts of our code run atomically (on one thread at a time). Critical section constructs use locks in order to achieve this mutual exclusion, but the actual implementation is hidden from us (abstraction). Similarly, atomic operations can mostly be translated into atomic instruction that are executed from the CPU.

Synchronization Hints

What if this pre-defined synchronization isn't enough for our needs? Well, its possible to define synchronization hints, which change the way the locks operate. Using hints its possible to specify if the critical section will be executed with low contention (few threads simultaneously) or high contention (many threads simultaneously). In the same way, it's also possible to specify if speculative techniques should be used. But, both basically affect only the speed of the synchronization.

The same synchronization hints can also be used in lock routines, that will be explained in a bit:

omp_lock_hint_none: the default behavior
omp_lock_hint_uncontented: low contention expected, which means that few threads are expected to perform the operation simultaneously
omp_lock_hint_contended: high contention expected, which means that many threads are expected to perform the operation simultaneously
omp_lock_hint_speculative: operation should be implemented using speculative techniques
omp_lock_hint_nonspeculative: operation should not be implemented using speculative techniques

Lock Routines

So, a critical section or atomic operation doesn't suit our needs - even with hints - what do we do? Well, don't worry, OpenMP offers Runtime Library Lock Routines for the management of Locks!

So, what is a lock? A lock (or mutex in some other APIs like POSIX), is an object that is used for mutual exclusion. Only the thread with the ownership of the lock may enter and execute the code inside of a critical section. A thread requests to acquire the ownership of the lock, before entering a critical section. And the thread inside the critical section releases the lock, right after it finishes executing the code inside the critical section. The lock has to be initialized before using it and destroyed after it isn't needed anymore.

In OpenMP a general-purpose lock can be of two types:

omp_lock_t: A simple lock that cannot be set again if it is already owned by the thread trying to set it
omp_nest_lock_t: A nestable lock that can be set multiple times by the same thread before being unset

OpenMP offers the following routines:

omp_init_lock(): Initialize a lock associated with the lock variablek
omp_init_lock_with_hint(): Initialize a lock associated with the lock variable using synchronization hints
omp_destroy_lock(): Disassociate the given lock variable from any locks
omp_set_lock(): Acquire the ownership of a lock, and suspend execution until the lock is set
omp_unset_lock(): Release a lock
omp_test_lock(): Attempt to set the lock, but don't block if the lock is unavailable

The same routines with nest_lock instead of lock in the name, are used for nestable locks.

Barriers In OpenMP

A barrier is a synchronization mechanism that makes all threads wait at the point of the barrier until all other threads have also reached the barrier. When no nowait clause is specified an implicit barrier is placed by OpenMP automatically at the end of each parallel region.

When the barrier has to be placed in a different section of the parallel code, explicitly, then a barrier directive has to be used.

The general format of such an directive is:

#pragma omp barrier

All threads will then wait until all other threads have reached that barrier, and then resume executing the parallel code that follows the barrier.

Achieving Memory Consistency: Flush Directive

There is one more directive that's quite useful in OpenMP, and that hasn't been covered yet, which is the flush directive. Using this directive its easy to specify a synchronization point at which the implementation must provide memory consistency. Even when the variables are shared, the threads might have their own point-of-view of memory. To optimize the code, the compiler might store those variables in different registers, making each thread have its own value. Thus, the flush directive has to be used, in order to ensure that the values observed by one thread are the same values that are observed by all the other threads.

The general format of the flush directive is:

#pragam omp flush [(list)]

The directive can optionally contain a list of variables that should be synchronized, so that only those are flushed. If no list is specified, then all variables that the thread can view are flushed.

And with that we basically covered all synchronization techniques that OpenMP offers!

Example Programs

So, after all that theoretical stuff, let's now get into some hands-on examples!

Synchronization using Locks

Let's start by implementing the code that we wrote in the Critical Sections article, about counting the zeros in a randomly-filled enormous array, using locks instead of the atomic operation.

Well, the only thing that has to change is that a lock has to be initialized before it's used inside of the parallel region, and destroyed (disassociated) after the region. To specify the counter incrementation as atomic the lock has to be set before the count++ statement and unset after it.

In C-code:

...

/* initialize lock */
omp_lock_t lock;
omp_init_lock(&lock);

    ...

    omp_set_lock(&lock);
    count++;
    omp_unset_lock(&lock);

    ...

/* destroy lock */
omp_destroy_lock(&lock);

...

Running this code we get the following console output:

Because the atomic operation on the CPU was replaced by locks, the performance got a little bit worse!

Synchronization using Barriers

To understand how the implicit barriers that OpenMP automatically defines and explicit barriers defined using the barrier directive operate, let's print out messages inside and outside parallel regions!

An Implicit Barrier is automatically defined at the end of each parallel region:

#pragma omp parallel
{
    printf("Thread %d running inside of the parallel region\n", omp_get_thread_num());
}

printf("Main thread after implicit barrier!\n");

Let's also print out multiple messages inside of the parallel region, to see the order in which the messages will be printed:

#pragma omp parallel
{
    printf("Thread %d running before explicit barrier\n", omp_get_thread_num());
    printf("Thread %d running after explicit barrier\n", omp_get_thread_num());
}

printf("Main thread after implicit barrier!\n");

An Explicit Barrier causes threads inside the parallel region to wait until all threads have reached the barrier, before they resume the execution of the parallel code:

#pragma omp parallel
{
    printf("Thread %d running before explicit barrier\n", omp_get_thread_num());

    #pragma omp barrier

    printf("Thread %d running after explicit barrier\n", omp_get_thread_num());
}

printf("Main thread after implicit and explicit barriers!\n");

The console output is:

As you can see the main thread executes after the parallel region has finished (implicit barrier), and the threads inside the parallel region with the explicit barrier wait for all threads to finish printing before going further!

Memory Consistency using Flush

Let's lastly also get into an example on how the flush directive might be useful.

Consider a for loop where a specific value is incremented by one thread (main thread), and the other threads printing out the value. The main thread has to use the flush directive after each incrementation, so that the other threads print out the correct value! Let's initialize a variable x to 0, and increment it by i for 10 iterations (1 to 10) of a for loop. The final value that we expect to see is 55.

The code for that is:

int x = 0;
int i;

#pragma omp parallel num_threads(8)
{
    if(omp_get_thread_num() == 0){
        for(i = 1; i <= 10; i++){
            x += i;
            #pragma omp flush
        }
    }
    else{
        sleep(1);
        printf("x : %d\n", x);
    }
}

If we haven't inserted the flush statement, some threads might print out 0 or other in-between values, even after the loop had finished, and even though the variable x is shared! The sleep statement was inserted just to make sure that the for loop finishes execution before-hand!

The output is:

It's also worth noting that the flush statement isn't always necessary as the implementation might do the job for us! The most optimized implementation of course doesn't specify different registers for each thread's variable in this case, but it might in other cases!

RESOURCES:

References

Images

https://commons.wikimedia.org/wiki/File:Fork_join.svg

Previous articles about the OpenMP API

OpenMP API Introduction → OpenMP API, Abstraction Benefits, Fork-Join Model, General Directive Format, Compilation
Parallel Regions in OpenMP → Parallel construct, Thread management, Basic Clauses, Example programs
Parallel For Loops in OpenMP → Parallel For Construct, Iteration Scheduling, Additional Clauses, Example programs
Parallel Sections in OpenMP → Parallel Sections Construct, Serial Section, Example Program
Atomic Operations and Critical Sections → Process Synchronization Theory, Critical Section Construct, Atomic Operations, Example Program
Tasks in OpenMP → Task Construct, Synchronization, Loops and Groups, Example Programs

Final words | Next up

And this is actually it for today's post!

Next time we will get into how we call SIMD Instructions from OpenMP...and I'm not sure if there is much more to cover after that!

Nvidia's CUDA API is already on the horizon :)

See ya!

Keep on drifting!

lock

Experts