Ensuring That Code in a Parallel Region Is Executed in Order

Ensuring That Code in a Parallel Region Is Executed in Order

In some cases, it may be necessary to ensure that a section of code is executed in the same order as the serial code would execute it. Unfortunately, such an ordering is unlikely to allow the code to get the full benefit of using multiple threads, but it should enable some gains to be attained from parallelization.

OpenMP supports the ordered directive, which ensures that the order of parallel execution is the same as the serial ordering. The directive needs to be applied to the par-allel region, and the loop also needs to be identified as an ordered loop using the ordered clause on the parallel for directive.

Listing 7.58 shows how the ordered directive can be used to ensure that the loop iterations are printed in the correct order.

Listing 7.58   Using the Ordered Directive to Ensure Code Executes in the Serial Order

#include <stdio.h> #include <omp.h>

int main()

{

#pragma omp parallel for ordered

for ( int i=0; i<100; i++ )

{

#pragma omp ordered

{

printf(" Iteration %i, thread ID %i\n", i, omp_get_thread_num() );

}

}

}

The ordered directive is most useful when applied to loops that do not use static scheduling. With the default static scheduling used in the example, the first thread will execute the first portion of the iterations, the second thread the second portion, and so on. Since the ordered region needs to be executed in the serial order, the second thread ends up waiting at the ordered code block until the first thread has completed all of its assigned work. This means that the work is serialized, but each serial chunk of work has been performed by a different thread.

The ordered directive is a useful way of exploring the impact of the scheduling on the order in which iterations are assigned to threads. Listing 7.59 shows the code modi-fied to use dynamic scheduling.

Listing 7.59  Using the Ordered Directive to Explore the Scheduling Directive

#include <stdio.h> #include <omp.h>

int main()

{

#pragma omp parallel for ordered schedule( dynamic ) for ( int i=0; i<100; i++ )

{

#pragma omp ordered

{

printf( "Iteration %i, thread ID %i\n", i, omp_get_thread_num() );

}

}

}

Listing 7.60 shows the effect of this change in scheduling. Dynamic scheduling causes the two threads to work with the default chunk size of a single iteration, so the two threads alternate performing iterations.

Listing 7.60   Exploring the Impact of Dynamic Scheduling

$ cc -O -xopenmp ordered.c $ export OMP_NUM_THREADS=2 $ ./a.out

Iteration 0 Thread 0

Iteration 1 Thread 1

Iteration 2 Thread 0

Iteration 3 Thread 1

...