MATLAB Distributed Computing Server and MATLAB job scheduler (MJS)

Hello,

I am very new to MATLAB. I am a sys admin and I have been ask to install and configure MATLAB Distributed Computing Server and MATLAB job scheduler (MJS).

I would like to:

Configure MATLAB Distributed Computing Server and MATLAB job scheduler (MJS) on 1 headnode and 7 machines. 224 cores.

Configure a MATLAB client machine with “Parallel Computing Toolbox”. For each member of the lab I would like to create their own cluster profile with 224 cores.

Members will login in the MATLAB client machine via Remote Desktop and run their codes.

I would like to know if userA login on the MATLAB client machine (via Remote Desktop) while userB’s job is running can userA using his/hers cluster profile queue his/hers job? In another word will userA's job run when userB's job is finished?

I hope I make sense.

Thank you

Best,

Eric

Answer 1

[Raymond Norris]

Any user will be able to start MATLAB on the client and submit a job, regardless if a job is already running. MJS is a fifo scheduler and if there are available workers to complete the job request, MJS will run conccurrent jobs (e.g., MJS could have 15-20 jobs queued by different users, while running 7-8 at the same time).

If this is a typical RDP (single user) and not a Terminal Server, then you could probably get away with a single profile, since only user will be logged into the RDP at a time. Be that as it may, if you create 1 or many profiles, I would suggest using Discover Clusters under the Parallel menu item on the Home toolstrip. From there, you can make copies of it, renaming each one of the profiles to be the names of your users.

[Masoud]

Thank you Raymond for your detailed answered. I watched several videos and that helped too. I will follow your advice and test it before deployment.

Would you build the MATLAB cluster on Linux or Windows? I am comfortable with either or.

Thank you,

Best,

Eric

[Raymond Norris]

Either will work fine, Eric. It might make the most sense though to keep the client and the compute nodes the same. This way if end users need OS specific files (i.e., MEX-files), they will transfer over fine. Also, they may have hardcoded paths (not ideal) that will transfer over easier (though not guaranteed) if it is at least the same OS.

[Masoud]

Great! Thank you very much Raymond. You have been a great help.

Best,

Eric