Relation between BLAS.get_num_threads() and SLURM_CPUS_PER_TASK

General Usage
,
1

I am running the script MWE.jl

using LinearAlgebra
display(BLAS.get_num_threads())

via SLURM command

srun --nodes=1 --ntasks-per-node=1 --cpus-per-task=6 --mem=1G julia MWE.jl

As far as I know, for SLURM the CPUs are logical, so if I set --cpus-per-task=6 I would expect that BLAS.get_num_threads() returns 6. Instead, it returns 3.

Why is that? Any help on this would be very much appreciated.

2

It’s on purpose that BLAS uses half of the available threads on x86_64 (to avoid potential hypethreading oversubscription), it’s also mentioned in Julia 1.12 Highlights. If you want to use a different number of OpenBLAS threads, set the env var OPENBLAS_NUM_THREADS appropriately.

3

I have done so via BLAS.set_num_threads and the impact on performance is terrible. Is that normal? The architecture of the computer is single-threaded. Even if it was multi-threaded, for SLURM the CPU number refers to logical CPUs, so I should always be able to set the same number of threads than logical CPUs.