Intel® Moderncode for Parallel Architectures
Support for developing parallel programming applications on Intel® Architecture.

Openmp task and parallel construct

Patrice_l_
Beginner
424 Views

Hi,

I am trying to understand the behavior of the Openmp implementation when a parallel do is enclosed in a task. When using nested  the parallel do uses multiple threads. The first question is is that possible to restrict the number of threads to the original thread pool (hardware thread), so that they work on the parallel construct has they become available after completing other task ? (see code below)

From reading the forum, i suspect the answer will be no, then what is the best way to combine task and parallel do , inside a task and outside a task.

Is it worth it to close the master or single region to do a parallel one, and reopen it right after ?

Last question, is there any  becnhmark of using task for a loop instead of a classic parallel do , in both case, fixed work load, and variable work load for each iteration ?

 

Thanks

program omptest
        use omp_lib
        implicit none
        integer :: i
!$omp parallel
!$omp master
print *,'omp get max threads',omp_get_max_threads(),omp_get_nested()
!$omp task
call omp_set_nested(.true.)
print *,'thread task1',omp_get_thread_num(),omp_get_max_threads(),omp_get_nested(),omp_get_thread_limit(),omp_get_num_threads()
!$omp parallel do
do i=1,10
print *,'thread task1 parallel do',omp_get_thread_num(),omp_get_ancestor_thread_num(omp_get_level()-1),omp_get_nested()
end do
!$omp end parallel do
!$omp end task
!$omp task
call omp_set_nested(.false.)
print *,'thread task2',omp_get_thread_num(),omp_get_max_threads(),omp_get_thread_limit(),omp_get_num_threads()
!$omp parallel do
do i=1,10
print *,'thread task2 parallel do',omp_get_thread_num()
end do
!$omp end parallel do
!$omp end task
!$omp taskwait
!$omp parallel do
do i=1,10
print *,'thread parallel do',omp_get_thread_num()
end do
!$omp end parallel do
!$omp end master
!$omp end parallel
end program omptest

 

0 Kudos
1 Reply
jimdempseyatthecove
Honored Contributor III
424 Views

If (IIF) your actual program is intended to be like your test program, then consider starting your outer parallel region with a limit of 2 threads. Then have each fork (assuming two), enable nested and set max active levels to 1, and have each set number of threads to half the threads. Then on your parallel for experiment with using dynamic scheduling.

Please report your findings.

Jim Dempsey

0 Kudos
Reply