Matlab Distributed Compute Server

The present version of Matlab is R2010b. The programs in the Matlab package are located in the directory /opt/matlab. Matlab production jobs are run on our so-called mini-cluster. The nodes of this cluster have between 8 and 24 cores and run under the Linux OS. Matlab processes can only be run in batch mode through our Gridengine scheduler.

Note that most Matlab Toolboxes are installed on the HPCVL clusters. However, HPCVL only holds a "3rd party access MATLAB Distributed Computing Server license". This enables HPCVL users who have a pre-existing Matlab license (Matlab + "Distributed Computing Toolbox (DCT)") to submit and run parallel Matlab jobs to our cluster nodes.

It is furthermore required that you sign a statement. We will confirm this statement, and you will then be made a member of a Unix group "matlab", which enables you to access the software. Contact us if you are in doubt of whether you will be able to run Matlab on our system.

Our Matlab license is "process limited". At present, there are 64 processes available. A single Matlab run can use up to 16 processes, and up to 2 Matlab jobs can run at a time for a given HPCVL user.

How do I run "Matlab" interactively?

The core of Matlab is a high-level programming language and an interactive programming/computing environment. Since you have to be a licensed user of Matlab to run Matlab programs on HPCVL machines, it is possible for you to either start a Matlab session on your own work station, or move the license to your HPCVL account and use our installation of the software. In either case, you are responsible for mainaining the license.

Here, we discuss how to run a Matlab session on HPCVL. The interactive part is usually started on a login node, for instance sflogin0. We assume that you are included in the "matlab" Unix group, and therefore have proper access to /opt/matlab where the software resides. To set the proper environment variables and include the directories with the binaries in your PATH, you should type:

use matlab

To avoid having to do this before every Matlab session, you may also include this command in your setup file, e.g. .bash_profile for a bash shell, or .login for a csh.

Note that for the sflogin0 login node, the "use matlab" command is setting you up for usage of the "R2009bSP1" version of Matlab. This is because that login node runs on the Solaris/Sparc platform and R2009bSP1 is the most recent version that is supported for that platform. This differs from the current version that is being used on the production nodes which run Linux (see next section). The above command is equivalent to adding the /opt/matlab/R2009bSP1/64bit/bin directory to your PATH. Finally you need to point the environment variable LM_LICENSE_FILE to the proper license file:

export LM_LICENSE_FILE=lic_file

for a bash shell. Here lic_file stands for the full path and file name of your license file. You need to make sure that the latter cannot be accessed by anyone but you. This command may of course also be placed into a setup file to avoid retyping. Once Matlab usage is set up properly, you only need to type

matlab

to start the GUI of the Matlab programming environment. Issuing the command with the -nodisplay will give you a command-line version that does not use the GUI.

In most cases, you will run Matlab in parallel "Batch" mode on our Linux "sw" cluster. Since you have to have access to Matlab outside of the HPCVL license, most interactive work could be done elsewhere, whereas the computationally intensive runs can be executed on HPCVL clusters.

Production jobs on HPCVL clusters are submitted via the GridEngine, which is a load-balancing software. To obtain details, read our Gridengine FAQ . For a Matlab batch job, this means that you have to prepare your Matlab setup to interface with GridEngine.

It is necessary for you to run the same version of Matlab on your local machine. This is currently R2010b. If you only have access to another version of the software, please contact us.

To setup running distributed jobs, you need to create a "Parallel Configuration" in Matlab. This needs to be done only once. Here are the steps involved:

• First, download a .zip file with several necessary files and extract the contents into the <MATLAB>/toolbox/local directory on your client machine. For Windows machines, the <MATLAB> portion of the directory path is something like C:\Program Files\MATLAB\R2010b.
• Open Matlab and enter the Parallel Configuration menu by clicking the Parallel->Manage Configurations tab. From there, select the File->Import... tab, and import the HPCVL-R2010b.mat from the <MATLAB>/toolbox/local directory. Now a new HPCVL-R2010b configuration should appear in the Configurations Manager window.
• Select the newly imported HPCVL-R2010b configuration, right click on it and open the "Properties" window for it. In the "SubmitFcn" and the "ParallelSubmitFcn" fields replace "hpc1708" by your username.
• You can test the configuration by pressing the "Start Validation" button in the lower right of the Configurations Manager window. You will be asked to provide login credentials for our machines. Type in your UserID and password, and answer the question if to use a login file with "No". If everything went alright, you should pass the "Find Resource" and the "Distributed Job" test. The "Parallel Job" test will fail, as we are currently unable to provide that functionality through our Gridengine.  

Once your parallel configuration is properly set up to enable job submission to the HPCVL production clusters, you should be able to submit distributed Matlab jobs from your Matlab instance.

Where can I get further help?