Parallel sever: Request GPU via SLURM

Hello,

I have access to an HPC and have been running CPU based code on it for some time now with help from the mathworks team. Parrellel server dir() for files local to the server - MATLAB Answers - MATLAB Central (mathworks.com)

I now wish to use GPU acceleration. When I try to use a gpuArray I get an error that the GPU devices are not found. This is because they were not requested in my batch command to SLURM which makes sense.

Is there a method to inject the additional SLURM gpu request argument e.g., --gres=gpu:2 into the .batch request (I am thinking something like the additionalSubmitArgs below) or is there a normal workflow for doing so without changing the submitCommunicatingJob function? It seems like a normal thing to request however I cannot find a straightforward workflow for it.

function job = submit_brainwave_job()
cluster = parcluster();

idir = '/inputDir'
idir = '/outDir'

job = cluster.batch(@batch_build_0_0_1,1,{idir,odir}, ...
    "AutoAddClientPath",false, "CaptureDiary",true, ...
    "CurrentFolder",".", "Pool",3);
%additionalSubmitArgs = sprintf('--ntasks=%d --gres=gpu:2', props.NumberOfTasks);
end

Kind regards,

Christopher

Answer 1

[Raymond Norris]

Hi @Christopher McCausland, try the following

cluster = parcluster();
cluster.AdditionalProperties.AdditionalSubmitArgs = '--gres=gpu:2';

You might also need to specify a gpu partition, as such

cluster.AdditionalProperties.AdditionalSubmitArgs = '--gres=gpu:2 -p <name-of-gpu-partition>';

Remember to unset AdditionalSubmitArgs if you don't need a GPU for you next job.

[Christopher McCausland]

Hi Raymond,

Thank you once again for the help! I will give this a go and let you know.

I take it any SLURM command request could be submitted via cluster.AdditionalProperties.AdditionalSubmitArgs = in the same way? For example;

cluster.AdditionalProperties.AdditionalSubmitArgs = '--gres=gpu:2','--cpus-per-gpu=2';

Matching the AdditionalSubmitArgs to the SLURM documentation on what can be requested?

Kind regards,

Christopher

[Raymond Norris]

Yes, that's right. In fact, for individual users familiar with their cluster, this is a viable option. When our team works with Cluster Admins, to make a more robust solution, we expand out AdditionalProperties, as such

cluster = parcluster();
cluster.AdditionalProperties.CpusPerGpu = 2;
cluster.AdditionalProperties.GpusPerNode = 2;
cluster.AdditionalProperties.Partition = 'gpu';

then on the backend, we implement getCommonSubmitArgs to convert this to the full Slurm syntax. This way users don't have to know the Slurm syntax themselves.

[Christopher McCausland]

Hi Raymond,

That is perfect, thank you for the help once again! This submission pipeline is brillent!

Kind regards,

Christopher