Limit CPU usage to one thread on cluster

Question

Hi,

I'm running a dolfin script on a cluster. I found out that this since job uses a CPU >> 100%, so it's not restricted to a single thread, which makes the running very inefficient and make the process claim way too much space on the cluster. I tried several things: I added

export OPENBLAS_NUM_THREADS=1

to my bash_profile and bashrc files. This works for non-dolfin code, but not for dolfin. I also tried something like

dolfin.parameters["num_threads"] = 1

following http://fenicsproject.org/qa/4078/dg-solve-in-parallel.
Both don't really help.
My question is: where to put the num_threads lines to make them effective? Is there a different way to limit CPU usage for dolfin jobs on a cluster to single thread per job?

Thanks,

Adriaan

Answer 1

If you set OMP_NUM_THREADS=1 in your environment variables, it should work for DOLFIN too.

Comments

I did add that to my bash files and restarted, but didn't help.

---

Are you sure the environment variables are applied when you start your parallel job? On most systems, setting OMP_NUM_THREADS=1 works. Maybe you can check your environment variables from within DOLFIN, e.g. by using os.environ

---

This helped me out! I added

import os
os.environ['OMP_NUM_THREADS'] = '1'

and all jobs are perfectly restricted to one thread.

Thanks a lot for the quick reply!