Parallel computing on a cluster
4 views (last 30 days)
Show older comments
Mads
on 7 Jul 2014
Commented: Johannes Kalliauer
on 3 Feb 2020
I have a script test.m that includes parfor-loops.
In MATLAB 2014a on my personal computer it runs the parallel job perfectly.
On a huge Linux computer cluster it runs test.m perfectly if I have started MATLAB 2014a graphically through X on the frontend.
However, when submitting test.m to the queue it discards the parallelness and runs everything as in for-loops -- on a single core on the given node.
What I write:
>> submat -q q12 test.m
q12 is the queue name.
Anyone with a clue??
2 Comments
Kevin Claytor
on 7 Jul 2014
This would probably be a question more suited for your cluster sysadmin. My guess is there's probably something in your queue submission file that is missing. Your cluster probably has a help page / wiki (for example, Duke's is: https://wiki.duke.edu/display/SCSC/DSCR ), I'd start there.
If you still can't find anything, we'll need some more details, for instance, what scheduler are you using? Is 'submat' a script or a command? If it's a script, can you post it?
Accepted Answer
Shashank Prasanna
on 7 Jul 2014
My guess is that when you queue it, it launches MATLAB without Java. Java is required to use PCT. That also explains why it works fine when you X11 forward MATLAB. See below:
"The client session of MATLAB must be running the Java® Virtual Machine (JVM™) to use Parallel Computing Toolbox software."
3 Comments
Shashank Prasanna
on 8 Jul 2014
you could alternatively consider "-nodesktop" which launches jvm but not the desktop.
Johannes Kalliauer
on 3 Feb 2020
use -noFigureWindows -nosplash -nodesktop -nodisplay instead of -nojvm.
More Answers (2)
Thomas Ibbotson
on 8 Jul 2014
We would need to see the code for 'submat', but my guess is that an independent job is being created rather than a communicating job. If you want to run a script with parfor loops on a cluster you need a communicating 'pool' job. For example you can submit one with 'batch' like this:
myCluster = parcluster('myClusterProfile');
job = batch(myCluster, 'test', 'Pool', myCluster.NumWorkers - 1);
wait(job);
fetchOutputs(job);
The 'Pool' argument instructs batch to create a communicating 'pool' job using the given number of workers to create the pool. You need to have at least 1 spare worker to act as the 'client', which is why I subtracted 1 from the total number of workers that the cluster has.
0 Comments
See Also
Categories
Find more on Parallel Computing Fundamentals in Help Center and File Exchange
Products
Community Treasure Hunt
Find the treasures in MATLAB Central and discover how the community can help you!
Start Hunting!