OpenMP threads executing on same CPU core

Question

Config : Windows 10, mingw-x64, Core 2 (4 logical)

The following code

#include <cstdio>
#include <omp.h>
#include <windows.h>

int main() {

    #pragma omp parallel
    {
        std::printf ("Doing Stuff on thread: %d  on Core #%d\n", omp_get_thread_num(),\
                      GetCurrentProcessorNumber());
    }

    return 0;
}

Sometimes it gives output:

Doing Stuff on thread: 1  on Core #1
Doing Stuff on thread: 0  on Core #2
Doing Stuff on thread: 2  on Core #2
Doing Stuff on thread: 3  on Core #2

Although most of the times it gives

Doing Stuff on thread: 2  on Core #1
Doing Stuff on thread: 1  on Core #3
Doing Stuff on thread: 0  on Core #2
Doing Stuff on thread: 3  on Core #0

which is expected.

How can I make sure that the threads have their affinity set for a different core? Is there a way to force the scheduler to do this?

@appleapple We are supposed to demonstrate parallel algo for coursework. Just don't want to make it run sequential just in case. Although I could run in linux if needed, wanted to know for windows too. — Recoder, Aug 13 '18 at 13:22
even if it runs on same core, it may run in parallel (interleaved). otherwise you cannot have more threads than core. — apple apple, Aug 13 '18 at 13:24
Since my laptop is dual core + hyperthreaded, I am not entirely sure whether the proc ID 2, 3 makes much of a difference. — Recoder, Aug 13 '18 at 13:31
Related question: https://stackoverflow.com/questions/663958/how-to-control-which-core-a-process-runs-on — Evg, Aug 13 '18 at 13:43
run in sequential (randomly) **is** what would really happen. avoid it is not good for learning parallel processing, I think. — apple apple, Aug 13 '18 at 13:54
you can `sleep` and force some thread finish later, it may be a better demo. — apple apple, Aug 13 '18 at 13:57
HW threads are not cores so forcing the affinity on the them might not be a good idea anyway. You should restrict your number of OpenMP threads to the number of core (2 here) with `OMP_NUM_THREADS=2`. Now, if you want also to force affinity, you can set `OMP_PROC_BIND=true` and `OMP_PLACES=cores` — Gilles, Aug 13 '18 at 14:21
Guessing you may be using gcc, libgomp on Windows doesn't support omp_places so the placement of threads will get vary. Although Windows was supposed to have hyperthreads scheduled since win7sp1 it doesn't keep 2 threads on separate cores and best performance is likely for 3 threads unless you can disable hyperthreads. — tim18, Aug 14 '18 at 13:22

OpenMP threads executing on same CPU core

0 Answers0