Saturday, June 20, 2009

OpenMP jobs on Grid? (The LCG-CE - PBS approach)

There was a user support requirement for OpenMP jobs in Grid. OpenMP is a shared-memory implementation which means that all processes must run on the same box.

Well this can easily achieved at PBS side by using the directive:
#PBS -l nodes=1:ppn=X

Where "X" is the number of requested processes. But the main issue is HOW can we get this requirement based on what WMS gives to us on submission?

After googling this, the "correct" solution can only be achieved at CREAM CE where users can select a number of requirements that will not only be used for job matching process at WMS but also passed to the CE. You can find more info on this here.

LCG CEs on the other hand are only getting a poor RSL which doesn't carry almost any of the user's requirements. So lets get in LCG CE's internals...

First a job reaches the globus-gatekeeper. At this phase user's proxy is matched to a pool account. GateKeeper's task is to authenticate the user and the job and pass it to the globus job manager.

The globus job manager uses the GRAM protocol to report the job state and submits the job to the globus-job-manager-marshal which is using a perl module to talk to the relevant queuing system.

This perl module is responsible for the creation of the job (shell script) that will be submitted to the PBS server. In this module the CpuNumber requirement is translated by default to:
#PBS -l nodes=X

So this is the part we need to change in order to create OpenMP jobs. The next issue now is how we find out if user has asked for OpenMP job. I've noticed that the JDL option "Environment" is passed to the job executable that will be submitted thus a definition like the following:
Environment = {"OPENMP=true"};
can do the trick.

The whole above approach works but for sure needs a lot of work but as proof of concept is more than ok...
In the (near) future i would like to test the CREAM CE which, as i said before, has a more clear way to support requirements from JDLs using the CeForwardParameters definition.

No comments:

Post a Comment