[EMAIL PROTECTED] schrieb am 05/05/2008 05:08:16 PM: > Hi, > > We want to execute a script in a cluster job submission via globusrun-ws, > but it should run exactly once, after stage-in, before processes are > started on the compute nodes. > > The purpose is to do some pre-run housekeeping in the user's home > directory. > > We thought we could do this by submitting a small script in the JDD > executable section, but it now appears that (for MPI jobs anyway) this is > executed on each compute node (as if it were the argument to mpirun in a > conventional cluster job submission). > > So we are in a sense trying to put something between the Job Description > document and the computation processes. > > This seems like a very natural thing to do: in conventional batch systems > it is done in a batch script. One can think of messy solutions where > the script detects whether it has been run before.... but that can't be > the right way. > > What is the right way to do this in a Globus job submission?
I agree that what you want is very natural. I think there is no right way just yet, see my report at http://bugzilla.mcs.anl.gov/globus/show_bug.cgi?id=5698 Meanwhile, we have implemented a "messy solution" in our project: https://bi.offis.de/wisent/tiki-index.php?page=Condor-GT4-BigJobs We also have an updated version which relies on a UDP broadcast for interprocess synchronization rather than file locking. It improves startup time compared to the NFS-based implementation, especially when you have many processes. However, it's not published yet. Let me know if you are interested in trying it out. Regards, Jan Ploski -- Dipl.-Inform. (FH) Jan Ploski OFFIS Betriebliches Informationsmanagement Escherweg 2 - 26121 Oldenburg - Germany Fon: +49 441 9722 - 184 Fax: +49 441 9722 - 202 E-Mail: [EMAIL PROTECTED] - URL: http://www.offis.de
