This would be easier (cleaner?) to accomplish with a SUBMITFILTER
script. This script, if defined in torque.cfg, takes the submitted
script on stdin and writes a modified script on stdout (which is then
used for the job). The script is completely general and can modify the
script anyway you'd like. See the MPI page on the LCG wiki for an
example of how to configure it. The MPI example actually modifies the
node specification, so is similar to what you want.
Cheers.
Cal
Steve Traylen wrote:
> On Tue, Mar 15, 2005 at 11:43:04AM +0100 or thereabouts, Manuel Sanchez
> Garcia wrote:
> > Hi,
> >
> > I'm trying to setup a torque based CE to use some nodes for a given
> > queue (lhcb) and different nodes for another queue (test). For this
> > purpose, I've assigned properties to the nodes (lcgpro and test) and
> > then included in the queue definition a requirement like:
> >
> > set queue lhcb resources_default.nodes = 1:lcgpro
> > set queue test resources_default.nodes = 1:test
> >
> > This works in the sense that if a job is submitted directly with qsub it
> > gets nodes=1:lcgpro as a requirement and then it is sent to the
> > appropiate nodes. So far so good, the problem is that it does _not_ work
> > for jobs submitted through LCG, they only get nodes=1 as requirement, as
> > reported with qstat -f.
> >
> > Looks like the default requirements declared in the queues are overriden
> > by the job manager. I've tried following the jobmanager scripts:
> > /opt/globus/lib/perl/Globus/GRAM/JobManager/pbs.pm
> > /opt/globus/lib/perl/Globus/GRAM/JobManager/lcgpbs.pm
>
> We have just been editing the lcgpbs.pm or rather the lcgpbs.in file
> that generates lcgpbs.pm so that jobs are submitted with
> qsub -l nodes=1:lcgpro
>
> You will see by default it is submitted with
>
> qsub -l nodes:1
>
> which overrides the default you have above.
>
> Steve
>
> >
> > But didn't see where the problem was. So the question is: am I
> > overlooking something? Did anybody see this behaviour before? Any clue
> > will be appreciated...
> >
> > Thanks in advance,
> > Manuel
>
> --
> Steve Traylen
> [log in to unmask]
> http://www.gridpp.ac.uk/
>
|