Matlab parfor on Slurm

Hi All,

I want to run a parfor on matlab using Slurm job submission on the cluster.

I want to use 15 workers for the parfor. Each worker should calls an external executable program and runs it using a set of distinc parameters. The external program needs 1 CPU(core) for each run.

I know that inside matlab m file I should set the parpool to 15. My question is that which of these two sbach job submission configurations I should to submit the job:

#SBATCH --ntasks=15
#SBATCH --cpus-per-task=1

or

#SBATCH --ntasks=1
#SBATCH --cpus-per-task=15 

[Pranav Nadigapu Suresh]

Hi,

Did you find answer to this?

Thanks,

Pranav

[H R]

Based on my trials I think option 2 would work. Although in some slurm user webpages they suggested option1, but the job submission was failed when I tried that

[Raymond Norris]

Hi,

In my experience, I would choose option #2, but keep in mind that the MATLAB client requires a CPU as well, so I would suggest --cpus-per-task=16.

However, you're also saying that each worker will call an external executable. Can the worker an executable it calls run on the same core? If not, perhaps you need 31?

1 (MATLAB) + 15 (workers) + 15 (executable program)

Raymond

Answer 1

[Jason Ross]

Note that you can specify these options as part of the ResourceTemplate property in the cluster (which would set this for every submission using the profile) or on the cluster object itself (which you could set on a per-job basis) for a Slurm cluster. See the "Slurm" section of the documentation here.

You can also set the Parallel Preferences to request 15 workers by default, or set the number of workers to 15 in the cluster profile.