openMP: Where to initialize the parallel region?

Question

I would like to run following code in parallel in openMP:

while (!stopcondition())
{       
 work_1();      
 work_2();  //work_2 has to be run after work_1
}

where the work() function is easy to parallelize:

work_i()
{
    //...some linear stuff...//

    #pragma omp parallel for schedule(static, 2) // make a piece of size 2 for each thread  
      for(int n=0; n<N; ++n)
         printf(" %d", omp_get_thread_num() );
}

It is inconvenient to set the parallel region inside the while loop or the work() function since it causes the threads to split each time, which results in big overhead.

When I tried:

#pragma omp parallel
 {
    #pragma omp single
    {
        while (!stopcondition())
        {
            work_1();       
            work_2();  
        }
    }
 }

Then both work_1 and work_2 always prints the same id, thus it is certainly executed by the same thread.

1) How can I overcome it? 2) Shall I just extract the // linear stuff // from work() and put it inside #pragma omp single?

3) Where is the best place to initialize the parallel region?

----------------------edit-----------------

It seems that the only way is to remove the stuff from work() and enclose it in a common parallel region:

   while (!stopcondition())
{
    #pragma omp parallel
    {
        #pragma omp parallel for schedule(static, 2) // make a piece of size 2 for each thread  
          for(int n=0; n<N; ++n)
             printf("Piece of work_1 %d", omp_get_thread_num() );
        #pragma omp parallel for schedule(static, 2) // make a piece of size 2 for each thread  
          for(int n=0; n<N; ++n)
             printf("Piece of work_2 %d", omp_get_thread_num() );
    }

    //...rest of linear stuff...//
}

The #pragma omp parallel set up a bunch of threads. Thanks #pragma omp single only one thread execute the code inside. — muaddieb
Generally it is not a good idea to use functions like printf in OpenMP parallel threads. — Trantor

jandres742 jandres742 · Accepted Answer · 2016-09-08T13:57:11

The problem with your last solution:

while (!stopcondition())
{
    #pragma omp parallel
    {
        #pragma omp parallel for schedule(static, 2) // make a piece of size 2     for each thread  
          for(int n=0; n<N; ++n)
             printf("Piece of work_1 %d", omp_get_thread_num() );
        #pragma omp parallel for schedule(static, 2) // make a piece of size 2     for each thread  
          for(int n=0; n<N; ++n)
             printf("Piece of work_2 %d", omp_get_thread_num() );
    }

//...rest of linear stuff...//
}

It's that every iteration of the loop will spawn threads, which is not good for performance. Moreover, you have nested parallel regions, which, depending on the OpenMP you are using, can create an oversubscription of threads: Each thread spawned by the first parallel will spawn a new set of threads for each inner parallel region; it's correct, but not optimal.

A solution would be:

#pragma omp parallel
{
  while (!stopcondition())
  {       
   work_1();      
   #pragma omp barrier /*So you make sure work_2 is executed after work_1*/
   work_2();  //work_2 has to be run after work_1
  }
}

Now, depending on what your stopcondition() function does, you might need some synchronization to make sure there are no race conditions and all threads exit the loop correctly.

openMP: Where to initialize the parallel region?

1 Answers