Is is possible to run a batch job on an MJS Cloud Center cluster with SpmdEnabled set to false?

I'm running some parallel pool computations on an MJS cluster (created with Cloud Center) using the batch command. Used this help article to set that up: https://www.mathworks.com/help/parallel-computing/run-a-batch-job.html#bu62o45.

I'm running with a pool of several hundred workers using the 'Pool' argument to the batch command. Unfortunately, the entire job will fail if any of the workers crash, which happens quite frequently.

Searching online, I've found that setting SpmdEnabled to false when using the parpool command will allow the task to complete on the remaining workers. I'd like to set this flag, but can't seem to find a way to do it using the batch command. Is there another way to disable SPMD support but also use the batch command with a parallel pool to submit a job to a cloud cluster?

Answer 1

[Edric Ellis]

Unfortunately, this option is not supported at the moment for batch jobs. I realise it's probably rather a big change to your code, but you could use independent tasks using createJob and createTask.

[Hridu Jain]

I've yet to test this, but can the SpmdEnabled flag be changed once the parallel pool is created? So, if the first lines of my function were to be: pool = gcp; pool.SpmdEnabled = false; Would that work?

[Edric Ellis]

No, you cannot set the SpmdEnabled flag after a pool has been constructed.

[Hridu Jain]

Thanks, Edric. If I were to use your suggestion and rework my code, would I need to eliminate use of parfor and parsim?

I believe parfor works when using job = createCommunicatingJob(...,'Type','pool',...) but this would once again have SPMD support, right?

So, I would instead need to rework the code to work with createJob and run createTask in a for loop. Is that right? And would there be anyway to use parsim with this setup?

Here's a rough outline of how my program is currently setup:

function y = run_simulations()

% initialize

[a,b] = initialize_stuff();

% preprocess and prep for simulation

parfor i = 1:500

SimIn(i) = preprocess_simulation_inputs(a,b);

end

% run simulations

SimOut = parsim(SimIn);

% post-process

y = postprocess(SimOut);

end

I call my run_simulations function with the batch command with the Pool argument.

[Edric Ellis]

I'm afraid you'd need to change your parfor loop into a series of independent tasks using createTask. (Likewise, you wouldn't be able to use parsim).