Matlab

This page will guide you through the serial computing with Matlab at Triton cluster. (Note (2017): We used to have the Matlab Distributed Computing Server (MDCS), but because of low use we no longer have a license. You can still run in parallel on one node, up to 20-28 cores depending on how new.)

Matlab configuration

Matlab writes session data, compiled code and additional toolboxes to ~/.matlab. This can quicky fill up your $HOME quota. To fix this we recommend that you replace the folder with a symlink that points to a directory in your working directory.

rsync -lrt ~/.matlab/ $WRKDIR/matlab-config/ && rm -r ~/.matlab
ln -sT $WRKDIR/matlab-config ~/.matlab
quotafix -gs --fix $WRKDIR/matlab-config

Interactive usage

Interactive usage is currently available via the sinteractive tool. Do not use the cluster front-end for doing heavy task. Only meant for submitting jobs/compiling. Using MDCS for sending jobs is ok.

ssh -X user@triton.aalto.fi
sinteractive
module load matlab
matlab &

Simple serial script

Running a single core Matlab job is easy through the slurm queue. A sample slurm script is provided underneath:

#!/bin/bash -l
#SBATCH -p short
#SBATCH -t 00:05:00
#SBATCH -n 1
#SBATCH --mem-per-cpu=100
#SBATCH -o serial_Matlab.out
module load matlab
n=3
m=2
srun matlab -nojvm -nosplash -r "serial_Matlab($n,$m) ; exit(0)"

The actual calculation is done in serial_Matlab.m-file:

function C = serial_Matlab(n,m)
        try
                A=0:(n*m-1);
                A=reshape(A,[2,3]).'

                B=2:(n*m+1);
                B=reshape(B,[2,3]).'

                C=0.5*ones(n,n)
                C=A*(B.') + 2.0*C
        catch error
                disp(getReport(error))
                exit(1)
        end
end

Remember to always set exit into your slurm script so that the program quits once the function serial_Matlab has finished. Using a try-catch-statement will allow your job to finish in case of any error within the program. If you don’t do this, Matlab will drop into interactive mode and do nothing while your cluster time wastes.

Multiple serial batchjobs

The most common way to utilize Matlab is to write a single .M-file that can be used to run tasks as a non-interactive batch job. These jobs are then submitted as independent tasks and when the heavy part is done, the results are collected for analysis. For these kinds of jobs the Slurm array jobs is the best choice; For more information on array jobs see Array jobs in the Triton user guide.

Below you will find an example how-to prepare and run such type of jobs.

run.m file doing the actual calculation task

The file below calculates Sin-function in the interval 0-2*PI and stores the results into a file. The interval is divided into blocks that are distributed over the nodes.

function run(blockIndex,pointsPerBlock,totalBlocks)
% blockindex runs from 0..totalblocks-1
% range 0..2pi
length=2*pi;
% values to setup even spacing between given range
% and splitting the spacings to even number of points per block
totalPoints=pointsPerBlock*totalBlocks;
step=length/(totalPoints-1);
start=blockIndex*pointsPerBlock*step;
% do some calculations, store the resulst so arrays A and B
for index=0:pointsPerBlock-1
  i=index+1;
  x=start+index*step;
  y=sin(x);
  A(i)=x;
  B(i)=y;
end
% save the results based on the blockIndex to a file
filename=strcat('output-',int2str(blockIndex));
save( filename, 'A', 'B', 'blockIndex');
% display message to output (log) that we have reached this far.
disp(sprintf('SUCCESS blockIndex %d',blockIndex));
% exit as this is a batch-job
exit;

Submission of 10 independent tasks

Below the run.m is executed as an array job with 10 array tasks, which will execute independently, potentially in parallel if there are enough idle resources. Note that it is using play partition with 5min time limit.

matslurm.sh:

#!/bin/bash -l
#SBATCH --time=0-00:05:00 --mem-per-cpu=500
#SBATCH -p debug
#SBATCH -o job-%a.out
#SBATCH --array=0-9
module load matlab
matlab -nojvm -r "run($SLURM_ARRAY_TASK_ID,100,10); quit"

Submit the job with “sbatch matslurm.sh” (or whatever you called the batch job script above).

Collecting the results

Finally a wrapper script to read in the .mat files and plots you tha Sin-function calculated in parallel with 10 tasks.:

function collectResults(numberOfBlocks)
   X=[];
   Y=[];
   for index=0:numberOfBlocks-1
      % read the output from the jobs
      filename = strcat( 'output-', int2str( index ) );
      load( filename );
      % catenate results to a single arrays
      X=cat(2,X,A);
      Y=cat(2,Y,B);
   end
   plot(X,Y,'b+:')

Seeding the random number generator

Note that by default MATLAB always initializes the random number generator with a constant value. Thus if you launch several matlab instances e.g. to calculate distinct ensembles, then you need to seed the random number generator such that it’s distinct for each instance. In order to do this, you can call the rng() function, passing the value of $SLURM_ARRAY_TASK_ID to it.

Parallel Matlab with Matlab’s internal parallelization

Matlab has internal parallelization that can be activated by requesting more than one cpu per task in the Slurm script and using the matlab_multithread to start the interpreter.

#!/bin/bash -l
#SBATCH -p short
#SBATCH -t 00:15:00
#SBATCH --nodes=1
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=4
#SBATCH --mem-per-cpu=2G
#SBATCH -o int_parallel.out

module load matlab
srun time -p matlab_multithread -nojvm -nosplash -r "int_parallel() ; exit(0)"

An example function is provided in this script

function int_parallel()
        try
                tic;
                A = rand(2000,2000);
                A = A + A.';
                B = pinv(A);
                max(max(B * A))
                toc
        catch error
                disp('Error occured');
                exit(0)
        end
end

Parallel Natlab with parpool

Often one uses Matlab’s parallel pool for parallelization. When using parpool one needs to specify the number of workers. This number should match the number of CPUs requested. parpool uses JVM so when launching the interpreter one needs to use -nodisplay instead of -nojvm. Example Slurm script:

#!/bin/bash -l
#SBATCH -p short
#SBATCH -t 00:15:00
#SBATCH --nodes=1
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=4
#SBATCH --mem-per-cpu=2G
#SBATCH -o parpool_parallel.out

module load matlab

srun matlab_multithread -nodisplay -r "parpool_parallel($SLURM_CPUS_PER_TASK) ; exit(0)"

An example function is provided in this script

function parpool_parallel(n)
        % Try-catch expression that quits the Matlab session if your code crashes
        try
                % Initialize the parallel pool
                c=parcluster();
                t=tempname()
                mkdir(t)
                c.JobStorageLocation=t;
                parpool(c,n);
                % The actual program calls from matlab's example.
                % The path for r2017b
                addpath(strcat(matlabroot, '/examples/distcomp/main'));
                % The path for r2016b
                % addpath(strcat(matlabroot, '/examples/distcomp'));

                % simulate 10000 blackjack hands with 100 players
                tic;
                pctdemo_aux_parforbench(10000,100,n);
                toc
        catch error
                getReport(error)
                disp('Error occured');
                exit(0)
        end
end

Parallel matlab in exclusive mode

#!/bin/bash -l
#SBATCH -p short
#SBATCH -t 00:15:00
#SBATCH --exclusive
#SBATCH -o parallel_Matlab3.out

export OMP_NUM_THREADS=$(nproc)

module load matlab/r2017b
srun -n 1 -c $OMP_NUM_THREADS matlab_multithread -nosplash -r "parallel_Matlab3($OMP_NUM_THREADS) ; exit(0)"

parallel_Matlab3.m:

function parallel_Matlab3(n)
        % Try-catch expression that quits the Matlab session if your code crashes
        try
                % Initialize the parallel pool
                c=parcluster();
                t=tempname()
                mkdir(t)
                c.JobStorageLocation=t;
                parpool(c,n);
                % The actual program calls from matlab's example.
                % The path for r2017b
                addpath(strcat(matlabroot, '/examples/distcomp/main'));
                % The path for r2016b
                % addpath(strcat(matlabroot, '/examples/distcomp'));
                pctdemo_aux_parforbench(10000,100,n);
        catch error
                getReport(error)
                disp('Error occured');
                exit(0)
        end
end

Hints for Condor users

The above example also works (even nicer way) for condor.

A wrapper script to execute matlab on the department workstation.

#!/bin/bash -l
# a wrapper to run Matlab with condor
block=$1
pointsPerBlock=10
totalBlocks=10
matlab -nojvm -r "run($block,$pointsPerBlock,$totalBlocks)"

Condor submission script

Condor actually contains ArrayJob functionality that makes the task easier.

## Condor submit description (script) file for my_program.exe.
## 1. Specify the [path and] name for the executable file...
Executable = run.sh
## 2. Specify Condor execution environment.
Universe = vanilla
notify   = Error
## 3. Specify remote execution machines running Linux (required)...
Requirements = ((OpSys == "Linux") || (OpSysName == "Ubuntu"))
## 4. Define input files and arguments
#Input = stdin.txt.$(Process)
Arguments = $(Process)
## 5. Define output/error/log files
Output = log/stdout.$(Process).txt
Error  = log/stderr.$(Process).txt
Log    = log/log.$(Process).txt
## 6. Tell Condor which files need to be transferred and when.
Transfer_input_files = run.m
Transfer_output_files = output-$(Process).mat
Transfer_executable = true
Should_transfer_files = YES
When_to_transfer_output = ON_EXIT
## 7. Add 10 copies of the job to the queue
Queue 10

FAQ / troubleshooting

If things randomly don’t work, you can try removing or moving either the ~/.matlab directory or ~/.matlab/Rxxxxy directory to see if it’s caused by configuration.

Random error messages about things not loading and/or something (Matlab Live Editor maybe) doesn’t work: ls *.m, do you have any unexpected files like pathdef.m in there? Remove them.

Also, check your home quota. Often .matlab gets large and fills up your home directory. Check the answer at the very top of the page, under “Matlab Configuration”.