Document requirement to set OMP_NUM_THREADS=1 when running bilby.run_sampler() under HTCondor
I spoke today with @gregory.ashton regarding the 'manual' equivalent of Abuse of CPU resources (bilby_pipe#206 - closed). I have a user on the cardiff cluster who is requesting a 16-core slot from HTCondor, then inside that slot running a Python script that executes bilby.run_sampler(..., npool=16, ...)
. However, because HTCondor defaults the job environment to include OMP_NUM_THREADS=$(RequestCpus)
, each process in the pool is attempting to use 16 threads, so the total CPU usage is well above the original request of 16.
Is it reasonable to document that when running bilby.run_sampler()
under HTCondor, users should be (explicitly) setting OMP_NUM_THREADS=1
in the environment?