Running Matlab Analyses

The typical procedure for performing calculations on UArizona HPC systems is to run your program non-interactively on compute nodes. The easiest way to run Matlab non-interactively is to use input/output redirection. This method uses Linux operators < and > to point Matlab to the input file and tell where to write the output (see the example script). The other method is to invoke Matlab from the SLURM script and execute specified statement using -r option. For details, refer to the manual page for the matlab command.

#!/bin/bash
#SBATCH --job-name=matlab
#SBATCH --account=group_name
#SBATCH --partition=standard
#SBATCH --ntasks=20 
#SBATCH --nodes
#SBATCH --mem-per-cpu=5gb
#SBATCH --time=01:00:00
 
module load matlab/<version>
 
matlab -nodisplay -nosplash < script_name.m > output.txt

The options -nodisplay and -nosplash in the example prevent Matlab from opening graphical elements. To view the full list of options for the matlab command, load the Matlab module and type matlab -h at the prompt. Alternatively, use the link above to see the manual page on the MathWorks website.

Parallel Computing Toolbox

Temporary Files

By default, Matlab PCT will dump files to ~/.matlab/MATLAB_VERSION. This causes problems when multiple Matlab PCT jobs are running simultaneously. Users should always define the environment variable MATLAB_PREFDIR so each job uses a unique temporary folder. Files there will be cleaned after the job finishes. For example:

export MATLAB_PREFDIR=$(mktemp -d $SLURM_JOBTMP/matlab-XXXX)

Matlab and SLURM Resource Requests

If you are trying to run Matlab in parallel interactively, you may encounter the following error:

>> Starting parallel pool (parpool) using the 'local' profile ...
Error using parpool (line 149)
You requested a minimum of <n> workers, but the cluster "local" has the NumWorkers property set to allow a maximum of 1 workers. To run a communicating job on more workers than this
(up to a maximum of 512 for the Local cluster), increase the value of the NumWorkers property for the cluster. The default value of NumWorkers for a Local cluster is the number of
physical cores on the local machine.

This is caused by an interaction between SLURM and Matlab. To resolve this issue, when requesting <n> cores for your interactive job, you will need to set SLURM's --ntasks directive to 1 and --cpus-per-task to the number of cores you need. For example:

$ salloc --nodes=1 --ntasks=1 --cpus-per-task=6 --mem-per-cpu=5GB --time=01:00:00 --job-name=interactive --account=<GROUP> --partition=standard

External Resources

If you are getting started with Matlab or think there might be a better way, check out the training resources.