Setting OMP_NUM_THREADS from octave

[François-Henry Rouet]

Hi all,

I'm trying to understand how octave deals with multithreaded libraries. I use octave 4.0.0, built with gcc 5.1.0, and linked against multithreaded OpenBLAS 0.2.14 (built with USE_OPENMP=1), which provides BLAS routines (e.g., matrix multiply) and LAPACK routines (e.g. LU factorization).

I wrote two mex functions to play with OMP_NUM_THREADS (wrappers to omp_[gs]et_num_threads):

void mexFunction(int nlhs, mxArray* plhs[], int nrhs, const mxArray* prhs[]) {
#pragma omp parallel
  printf("OMP_NUM_THREADS=%d\n",omp_get_num_threads());
}

void mexFunction(int nlhs, mxArray* plhs[], int nrhs, const mxArray* prhs[]) {
  double *dn = mxGetPr(prhs[0]);
  omp_set_num_threads((int)(dn[0]));
}

Now here's the issue; in this example, I start octave with OMP_NUM_THREADS undefined. I set the number of threads to 2 with my mex function, run some BLAS/LAPACK functions, then I set the number of threads to 1, run some BLAS/LAPACK functions, then I set the number of threads to 2 again.

>> omp_set_num_threads(2)
>> omp_get_num_threads()
OMP_NUM_THREADS=2
OMP_NUM_THREADS=2
>> n=6000; A=rand(n,n)+n*eye(n); tic; lu(A); toc; tic; A*A; toc;
Elapsed time is 8.04974 seconds.
Elapsed time is 16.3446 seconds.
>> omp_set_num_threads(1)
>> omp_get_num_threads()
OMP_NUM_THREADS=1
>> n=6000; A=rand(n,n)+n*eye(n); tic; lu(A); toc; tic; A*A; toc;
Elapsed time is 10.5778 seconds.
Elapsed time is 30.6964 seconds.
>> omp_set_num_threads(2)
>> omp_get_num_threads()
OMP_NUM_THREADS=2
OMP_NUM_THREADS=2
>> n=6000; A=rand(n,n)+n*eye(n); tic; lu(A); toc; tic; A*A; toc;
Elapsed time is 10.5794 seconds.
Elapsed time is 31.3423 seconds.

Going from 2 to 1 thread seems to work. However, when i set the number of threads back to 2, I still get single-thread performance. Also, if I start octave with OMP_NUM_THREADS=1, I get single-threaded behavior regardless of my calls to omp_set_num_threads.

Any explanation to this? I found some old messages in the matlab and octave archives but nothing really helpful.

Also, I tried to use setenv/putenv to set OMP_NUM_THREADS and it doesn't seem to work at all.

Thanks for your help!