Running compiled Matlab under Condor

Problems

Multiple users on the same node

When running compiled Matlab codes under Condor, you need to be aware that the default settings of Matlab as well the Matlab runtime environment (also called MCR) might bite you. It seems that when a MCR job is run, the directories /tmp/.matlab and /tmp/.mcr_cache_v78 (for Matlab 2008a) are being created and used by the executables. However, these are most of the times not cleaned up after the job is finished and if another user's MCR program wants to start on this node, it will wait forever since it cannot create a lock file in this directory as it is owned by another user. It is currently not fully clear if this could also be an issue if the same user runs multiple different programs on the same nodes. Solution...

Matlab runs on all available cores

Matlab executable starts on all available cores, which does not work well with condor. Solution

Possible solutions

Multiple users on the same node

After consulting with Keith Thorne and Jamie Rollins the attached wrapper script was devised and should solve this issue by creating a temporary unique directory and redirecting $HOME to that place. Please report success and failures to us!

It is important to use the wrapper script as an executable, e.g. if your mcc-compiled binary is named GW_detection and you need the arguments a 50 --verbose you need to run it like

/path/to/atlas-matlab-init /path/to/GW-detection a 50 --verbose

or in your submit file:

Executable     = /path/to/atlas-matlab-init
Arguments      = /path/to/GW_detection a 50 --verbose
[...]

Matlab runs on all available cores

Either use File->Preferences->General->Multithreading from the Matlab menu before compiling your code or add maxNumCompThreads(1) to you script.

-- CarstenAulbert - 10 Jun 2009