What is the Matlab Distributed Computing Server?

  • The MATLAB Distributed Computing Server (MDCS) allows users to submit  (from within MATLAB) sequential or parallel MATLAB code to a cluster.
  • After submitting a job, a user can continue using MATLAB, submit more jobs, or turn off their computer; Once the submitted job(s) have completed, the data can be retrieved from the user's storage space on a cluster.
  • Under the current CCR MDCS purchase, MATLAB can be used on a maximum of 256 cores at once.

General requirements for use

  • CCR account
  • We recommenced using the most recent version of MATLAB installed on the UB CCR.  A personal desktop installation of MATLAB is not required.

Getting Started

The tutorial below will compute Pi using a Monte Carlo method. An additional MDCS example is available on the vortex front-end under "/util/academic/matlab/example" ---- the files "My_MDCS_Script.m" and "mbatchSlurm.m" contain an example of running a simple MDCS job on the CCR cluster.

Step 1:   Connect to vortex.ccr.buffalo.edu and create a directory for temporary storage of MATLAB output. We will refer to this as StorageLocation. Create this directory under /panasas/scratch/grp-YourGroup (replace with your group's scratch directory), which has a 10TB storage quota.  Note, that all files in /panasas/scratch that have not been accessed within 30 days are automatically deleted - be sure to copy important files out of this directory before they are deleted.


           [user@vortex]$ cd /panasas/scratch/grp-YourGroup

           [user@vortex]$ mkdir usernameMatlabData


Step 2: Copy the files mbatchSlurm.m and piMC.m from /util/academic/matlab-scripts to your desired working directory.

           [user@vortex]$ cd ~

           [user@vortex]$ mkdir MatlabTutorial
           [user@vortex]$ cd MatlabTutorial
           [user@vortex]$ cp /util/academic/matlab-scripts/* ./

Step 3: Load the MATLAB module, start matlab, and edit the script mbatchSlurm.m

           [user@vortex]$ module load matlab
           [user@vortex]$ matlab -nodesktop
              < M A T L A B (R) >
              Copyright 1984-2014 The MathWorks, Inc.
            >> edit mbatchSlurm.m  

Step 4: Edit the submit-script. The MATLAB submit-script is named mbatchSlurm.m and it contains the key information that MATLAB needs to run your code on the cluster. The following lines should be modified:

  • Specify chosen StorageLocation:
        StorageLocation='/panasas/scratch/grp-YourGroup/usernameMatlabData';
  • Request the number of processors-per-node (ppn) for your job.
        ppn=12 
  • Request walltime that is expected to exceed that of the actual computation. The maximum CCR walltime is 72 hours (72:00:00) Example - for 10 minutes:
        time='00:10:00';
  • Specify an e-mail address where notification of job completion will be sent.
    email='username@buffalo.edu';
  • Any additional SLURM directives that you would like to include for your job should also be included in the e-mail string. For example, to request a CPU-X5650 processor type that has at least 48GB RAM per node and two GPUs:
        email='username@buffalo.edu --constraint=CPU-X5650 --gres=gpu:2 --mem=48000';
  • Specify the number of workers. Each worker is associated with a labindex (rank, process id). For example:
        pjob.NumWorkersRange = [1 24];

     Initializes the parallel job 'pjob' to run over 24 "matlabs" (workers)

  • List all files need to run your program (ex. main code, your functions, input data, etc.).  First try running the example code  piMC.m.  Note: Your main code must be written as function, with an input and an output. The input and ouptut to this function can be made something trivial, like a print statement.   Your actual output data can be saved as a MAT file from within your main code.
        set(pjob, 'AttachedFiles',{'piMC.m'});

    • NOTE: attaching large (>100 MB) data files can cause problems for MDCS. As a workaround, instead of attaching large files you can just add the containing folder to your path using the "AdditionalPaths" option:
          set(pjob, 'AdditionalPaths', {'/path/to/data'});
  • Create a task to run over the workers.   There are 3 fields here. In order, they are the main code, the number of outputs from the main code, and the value(s) of the input(s) to the main code.

        createTask(pjob,@piMC,1,{5e6});

Step 5: Submit the mbatchSlurm.m job.When you are ready to submit your job, you will simply run the script from within MATLAB. 

        >> mbatchSlurm

If everything is working correctly, the prompt should return in less than a minute or two. You can now check the status of the job:

        >> ! squeue --user=your_user_name

You should notice your job in the queue with the job id assigned by MATLAB as the job-name.  Also, a folder with this job name should be created in StorageLocation. Once the job has finished, the file stuff.mat should be created in StorageLocation/Job#.  Navigate to this folder and extract the results into MATLAB

        >> cd /panasas/scratch/grp-YourGroup/usernameMatlabData/Job#

        >> load stuff.mat
        >> mypi

If a value returns, then great. You just computed an approximation to pi using the MDCS.  Now try it with your own code!