When submitting jobs to a third party scheduler, if it's necessary to pass additional arguments to the scheduler, rather than modifying the configuration, properties can be stored using the singleton class, ClusterInfo.
For example, if the Sys Admin wants the user to specify the walltime of a job, the user can specify it as such:
>> ClusterInfo.setWallTime('04:00:00');
>> job = batch(@simulation,1,{2,4});
In turn, the integration script used by the configuration will query for the walltime as such:
wt = ClusterInfo.getWallTime();
if isempty(wt)
% Not set, default to 10 minutes
wt = '00:10:00';
end
pn = ClusterInfo.getProjectName();
if isempty(pn)
% Not set, but required, so error out
error('For billing purposes, you must specify the project name')
end
...
% Embedded user options to call
cmd = system(['qsub -A ' pn ' -l walltime=' wt ' ...]);
The properties that can be passed are:
% Arch
% ClusterHost
% EmailAddress
% MemUsage
% ProcsPerNode
% ProjectName
% QueueName
% Reservation
% UseGpu
% UserDefinedOptions
% UserNameOnCluster
% WallTime
If an additional property is required, the Admin should create a setter and getter for it in ClusterInfo.m. Alternatively, the user can pass an argument to 'UserDefinedOptions', for instance
>> ClusterInfo.setUserDefinedOptions('-l pvmem=2gb')
The advantage to using ClusterInfo is that it allows the Admin to support a single configuration for their cluster while allowing the user the ability to pass additional arguments, tailoring it to their specific job submission.
Once the property is set, it is persistent between jobs as well as MATLAB sessions for that user. Call ClusterInfo.state to get the state of each property and ClusterInfo.clear to clear all property values.
ClusterInfo can be used by any integration script (i.e. any scheduler).
|
