MATLAB on Ganymede

MATLAB jobs can be run in serial and parallel modes, in the background with sbatch or interactively with srun or salloc on CIRC HPC systems, except Europa.

For more information about the interactive jobs using srun or salloc, please refer interacting jobs.

MATLAB in Interactive Mode

To run MATLAB interactively from a Ganymede cluster node, log in to a compute node, then load the MATLAB module:

srun -N 1 -n 1 --cpus-per-task=16 --pty /bin/bash
or
salloc -p debug -N 1 -n 4 --time=00:30:00
module load matlab
matlab -nodisplay -nosplash

Do not run MATLAB in the login node. This may result in the system running out of resources and impacting other users.

Execution will take some time, normally a few moments. Now, the user can execute MATLAB functions interactively.

[@compute-7-6-0 ~]$ ml matlab/R2020b
[@compute-7-6-0 ~]$ matlab -nodisplay -nosplash
MATLAB is selecting SOFTWARE OPENGL rendering.

                                                                < M A T L A B (R) >
                                                      Copyright 1984-2020 The MathWorks, Inc.
                                                  R2020b Update 1 (9.9.0.1495850) 64-bit (glnxa64)
                                                                 September 30, 2020


To get started, type doc.
For product information, visit www.mathworks.com.

>>
>> ver				# For a complete list of MATLAB, Simulink, and MATLAB Toolboxes and
			        # their version information, type the "ver" command at the MATLAB prompt

When finished, exit from the MATLAB terminal and the compute node.

>> exit
[0@compute-7-6-0 ~]$ exit
exit
[@ganymede ~]$

Running MATLAB Batch Jobs

Create a new folder called matlab in your scratch directory.

mkdir matlab && cd matlab

Copy the submission script matlab-job-script and MATLAB file matlabTest.m from the directory /opt/ohpc/pub/examples/matlab/ to $SCRATCH . The submission script should be done as shown in an example script in the directory /opt/ohpc/ pub/examples/matlab/matlab-job-script.

#!/bin/bash
#SBATCH --partition=normal
#SBATCH --job-name=Matlab-Test
#SBATCH --mail-user=`whoami`@utdallas.edu
#SBATCH --mail-type=ALL
#SBATCH --nodes=1
#SBATCH --ntasks=4
#SBATCH --output=r.result
#SBATCH --partition=normal
#SBATCH --account=`whoami`

module load matlab
matlab -nodisplay -nosplash < matlabTest.m > myResult

Running MATLAB Serial Jobs

For an example of a serial job please look in the /opt/ohpc/pub/examples/matlab/serial/ and follow the instructions in README file. Copy slurm-matlab-serial.sh and spectral_radius-serial.m to your scratch directory. Write a Slurm script as follows to submit:

#!/bin/bash
#SBATCH --job-name=sr-serial
#SBATCH --nodes=1
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=16
#SBATCH --output=sr-serial.out
#SBATCH --partition=normal
#SBATCH --mail-type=ALL
#SBATCH --mail-user=`whoami`

module load matlab
matlab -nodisplay -nosplash < spectral_radius-serial.m

Now submit the batch script with the sbatch command:

sbatch slurm-matlab-serial.sh

When run is complete check the output sr-serial.out.

$ cat sr-serial.out
MATLAB is selecting SOFTWARE OPENGL rendering.

                                                                < M A T L A B (R) >
                                                      Copyright 1984-2020 The MathWorks, Inc.
                                                  R2020b Update 1 (9.9.0.1495850) 64-bit (glnxa64)
                                                                 September 30, 2020


To get started, type doc.
For product information, visit www.mathworks.com.

>>
LASTIN =
16
>> >> >> >> >> >> Elasped time is 36.537003 seconds.

MATLAB Parallel Computing Toolkit (PCT) Jobs

Parallel Computing Toolkit (PCT) utilizes MathWorks' Parallel Computing Toolbox to parallelize MATLAB applications without MPI or CUDA programming. These jobs follow the directives of the Toolkit. MATLAB supports process parallelization and thread parallelization using parpool. Thread parallelization can be better when acting on a large, shared dataset.

parpool();              % Default parpool
parpool(16);            % Process parpool with 16 workers
parpool("threads");     % Threads parpool
delete(gcp('nocreate')) % deleting a parallel pool

Parpool(16) spawns 16 matlab processes or copies. Whereas threads in parpool("threads") uses single process and less time. Check using top -u <userid> command.

>> parpool(16)
Starting parallel pool (parpool) using the 'local' profile ...
Connected to the parallel pool (number of workers: 16).

ans =

 ProcessPool with properties:

            Connected: true
           NumWorkers: 16
              Cluster: local
        AttachedFiles: {}
    AutoAddClientPath: true
          IdleTimeout: 30 minutes (30 minutes remaining)
          SpmdEnabled: true

>>

>> delete(gcp('nocreate'))
Parallel pool using the 'local' profile is shutting down.
>> parpool("threads")
Starting parallel pool (parpool) ...
Connected to the parallel pool (number of workers: 40).

ans =

  ThreadPool with properties:

    NumWorkers: 40

>>

For an example of a parallel job, look in the directory /opt/ohpc/pub/examples/matlab/parallel/ and follow the README file. Copy both Slurm script slurm-matlab-parfor.sh

#!/bin/bash
#SBATCH --job-name=sr-parfor
#SBATCH --nodes=1
#SBATCH --ntasks=8
#SBATCH --output=sr-parfor.out
#SBATCH --partition=normal
module load matlab
matlab -nodisplay -nosplash < spectral_radius-parfor.m

and MATLAB script spectral_radius-parfor.m to your scratch directory. Now submit the job using sbatch command:

$ sbatch slurm-matlab-parfor.sh
Submitted batch job 600

When the run is complete, check output sr-parfor.out

$ cat sr-parfor.out
MATLAB is selecting SOFTWARE OPENGL rendering.

                                                                < M A T L A B (R) >
                                                      Copyright 1984-2020 The MathWorks, Inc.
                                                  R2020b Update 1 (9.9.0.1495850) 64-bit (glnxa64)
                                                                 September 30, 2020


To get started, type doc.
For product information, visit www.mathworks.com.

>>
LASTIN =
16
>> >> Starting parallel pool (parpool) using the 'local' profile ...
connected to 8 workers
>> >> >> >> >> >> Elasped time is 6.484026 seconds.

Example I : <matlabTest.m>

n=100;
for i=1:n
    disp(i);
end;

Example II : <spectral_radius-serial.m>

LASTN=maxNumCompThreads(1)
tic
n=200
A=500
a=zeros(n);
for i=1:n
    a(i)=max(abs(eig(rand(A))));
end
toc

Example III : LARGE ARRAY EXAMPLE: PROCESS PARALLELIZATION

parpool(16);
n = 1000000000;
a = zeros(1,n);
tic
parfor i = 1:n
    a(i) = -log(1 - rand());
end
toc

The output :

>> >> Starting parallel pool (parpool) using the 'local' profile ...
Connected to the parallel pool (number of workers: 16).
>> >> >> >> >> >> Elasped time is 15.580450 seconds.

Example IV : LARGE ARRAY EXAMPLE: THREAD PARALLELIZATION

parpool("threads");
n = 1000000000;
a = zeros(1,n);
tic
parfor i = 1:n
    a(i) = -log(1 - rand());
end
toc

The output:

>> >> Starting parallel pool (parpool) using the 'local' profile ...
Connected to the parallel pool (number of workers: 16).
>> >> >> >> >> >> Elasped time is 10.882831 seconds.