Matlab Parallel Server

Introduction

MATLAB Parallel Server allows you to run parallel jobs efficiently on Rockfish by using a cluster profile that integrates with SLURM. This approach avoids requesting a separate MATLAB license for each parpool or parfor call and improves resource utilization.

Create a Cluster Profile

You only need to create the cluster profile once.

Note

You can create the cluster profile by launching MATLAB either through Open OnDemand or from the command line. This tutorial uses the command-line method (without GUI).

Option: Create Profile via Command Line

Start an interactive session on a compute node:
```
interact -p shared -n 4 -t 01:00:00
```
Load the MATLAB module:
```
module load matlab/R2024a
```
Launch MATLAB in terminal mode:
```
matlab -nodisplay -nosplash
```
Inside MATLAB, create the cluster profile:
```
configCluster;
```
This creates a cluster profile named rockfish.
Exit MATLAB:
```
exit
```
Exit the interactive session:
```
exit
```

Write Your MATLAB Parallel Script

To run a parallel job, set up the cluster configuration and define the desired resources in your MATLAB script.

Example: test_matlab_parallel.m

% test_matlab_parallel.m

% Load the cluster profile
rf = parcluster('rockfish');

% Set SLURM resource parameters
rf.AdditionalProperties.Partition = 'parallel';  % Specify partition
rf.AdditionalProperties.WallTime = '02:00:00';   % Wall time must match SLURM script
rf.AdditionalProperties.AdditionalSubmitArgs = '--nodes=2 --ntasks-per-node=32 --cpus-per-task=1';

% Display properties (optional)
disp(rf.AdditionalProperties);

% Start parallel pool
disp('Starting a parallel pool...');
parpool(rf, 64);  % 2 nodes × 32 tasks

% Run a sample parallel computation
disp('Running parallel computation...');
results = zeros(1, 64);
parfor i = 1:64
    results(i) = i^2;
end
disp('Results:');
disp(results);

% Shut down the parallel pool
delete(gcp('nocreate'));

Submit Your Job to SLURM

While you can run MATLAB interactively, the recommended approach is to submit your job using a SLURM script.

Important

The --time specified in your SLURM script must match the WallTime defined in your MATLAB cluster profile. Other SLURM parameters in the script (e.g., --nodes, --ntasks) are not relevant because resource allocation is handled by MATLAB itself.

Example: run_matlab_test.sh

#!/bin/bash
#SBATCH --job-name=matlab_test
#SBATCH --nodes=1
#SBATCH --time=02:00:00      # Must match WallTime in MATLAB script
#SBATCH --partition=shared
#SBATCH --output=matlab_test.out
#SBATCH --error=matlab_test.err

module load matlab/R2024a

# Run MATLAB in batch mode
matlab -nodisplay -nosplash -r "test_matlab_parallel; exit"

Submit the SLURM Job:

sbatch run_matlab_test.sh

Final Notes

Only set up the cluster profile once per user.
Always ensure WallTime in the MATLAB script matches the SLURM --time value.
MATLAB handles resource allocation via the SLURM backend — no need to manually manage --nodes or --ntasks in the SLURM script.
You can view SLURM output and MATLAB results in the .out and .err files generated by SLURM.