Controlling the OpenMP Runtime Environment

Controlling the OpenMP Runtime Environment

The OpenMP runtime environment can be controlled in up to three different ways. We have already encountered the environment variable OMP_NUM_THREADS to set the num-ber of threads that the program uses. However, it is also possible to set this through pro-grammatic calls to the runtime library or even as clauses placed onto the directives in the source code. Clauses will override the settings from calls to API functions, and these will override any environment settings. This section discusses the various settings that can be configured and the options available for configuring them.

Setting the Number of Threads

As previously seen, the number of threads used by an OpenMP application can be set through the environment variable OMP_NUM_THREADS. It is also possible to set the number of threads using the function call omp_set_num_threads(), as shown in Listing 7.46. Calls to omp_set_num_threads() change the default value for all subsequent parallel regions. It is possible to determine the number of threads using the function call omp_get_max_threads(). The function call omp_get_thread_num() will return a unique ID for each thread.

Listing 7.46   Setting the Number of Threads

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

int main()

{

double total = 0.0; double array[1000];

omp_set_num_threads( 2 );

#pragma omp parallel for reduction( +: total ) for ( int i=0; i<1000; i++ )

{

total += array[i];

}

printf( "Total=%f\n", total );

printf( "Threads=%i\n", omp_get_max_threads() );

}

The number of threads for a parallel region can be specified in the source code using the num_threads(threads) clause. The value for the number of threads can be fixed, or it can be an integer calculated based on some other factors. Listing 7.47 shows an example of using a fixed value for this.

Listing 7.47   Setting the Number of Threads

#include <omp.h>

#include <stdio.h>

int main()

{

double total = 0.0;

double array[1000];

#pragma omp parallel for reduction( +: total ) num_threads( 2 ) for ( int i=0; i<1000; i++ )

{

total += array[i];

}

printf( "Total=%f\n", total);

printf( "Threads=%i\n", omp_get_max_threads() );

}

The num_threads clause will override the default value just for this single parallel region. The next parallel region will again take the default value for the number of threads, unless this too has a num_threads clause.

There is one other environment variable that can set the number of threads. OMP_THREAD_LIMIT sets the maximum number of threads that are allowed. It is an implementation defined as to whether this limit will be imposed on all attempts to use more threads than this limit. The value for this limit can be obtained through the func-tion call omp_get_thread_limit().

An OpenMP implementation can honor the environment variable OMP_DYNAMIC. This environment variable can be set to either true or false. If it is set to true, then the OpenMP implementation can react to runtime conditions and use fewer threads than requested for any parallel region. This variable can be set at runtime with a call to omp_set_dynamic(), and its value can be read by a call to omp_get_dynamic().

Setting Runtime Loop Scheduling

The scheduling for loops with the runtime scheduling clause is controlled with the envi-ronment variable OMP_SCHEDULE. The schedule can also be set at runtime through a call to omp_set_schedule(), and the current schedule can be obtained through a call to omp_get_schedule(). The function calls to get and set the schedule take two parame-ters. The first is an integer that indicates the scheduling requested. The second is the chunk size (the second parameter will be ignored for schedules that do not require a chunk size). The available schedules are omp_sched_static, omp_sched_dynamic, omp_sched_guided, and omp_sched_auto. Listing 7.48 demonstrates using the calls to get and set the schedule.

Listing 7.48   Getting and Setting the Schedule

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

int main()

{

omp_sched_t schedule; int chunksize;

omp_get_schedule( &schedule, &chunksize );

printf( "Schedule = %i, chunksize = %i\n", schedule, chunksize );

omp_set_schedule( omp_sched_guided, 10 );

}

Specifying the Stack Size for Worker Threads

The stack size of the master thread is set through the normal operating system environ-ment. This can be changed using the ulimit command on UNIX-like platforms. On Windows, the default stack size is set at link time.

The default stack size for each worker thread created by the OpenMP runtime library is implementation specific. Depending on the requirements for stack space, this default may not be sufficient. The environment variable OMP_STACKSIZE determines the stack space for the worker threads. There is no call into the runtime library that sets this size.

The environment variable takes a number with an optional suffix. A number with no suffix is interpreted as kilobytes, the suffix B indicates that the number is in bytes, the suffix K indicates that it is in kilobytes, the suffix M is interpreted as megabytes, and the suffix G indicates gigabytes.