Gordon, JC (John) wrote:
> Thanks Steve. This isn't really an operational matter for the
> Monday(Wednesday) meeting. It has been an SA1 requirement for at least a
> couple of years. At HEPiX last October, Francesco finally understood
> what was required and why and it has been on JRA1 workplan since then.
> Something was delivered but I never heard how the testing went.
The slides are available here:
http://hepix.caspur.it/spring2006/TALKS/4apr.prelz.dir/index.html
They're slightly depressing reading I'm afraid as they seem to have just
gone for the quick bodge option to shut the sysadmins up. For me there are
probably two separate issues that should probably be fixed separately:
1) The standard job submission parameters (walltime, cputime, etc.)
supported by Globus RSL and every LRMS under the sun _should_ be parsed
out from the JDL ClassAds on the WMS and the appropriate variables set in
the RSL. If we're using RSL as a layer of indirection we should take
advantage of it fully.
2) There is a potentially unlimited number of other useful parameters that
users could set in the requirements of the JDL. For example, you could
imagine that software requirements (e.g. MPICH) could be used by a local
batch system to load the appropriate software modules for the job and
potentially tweak the submission script accordingly (e.g. by doing an
mpiexec). For this there would need to be some agreement on naming of the
requirements.
To step back a bit, I find it depressing that no-one in JRA1 seems to be
thinking about these real-world issues: it applies across the board from
MPI support to passing batch systems requirements to short jobs etc. etc.
Is there anything we can do to improve this?
Stephen
--
Dr. Stephen Childs,
Research Fellow, EGEE Project, phone: +353-1-8961797
Computer Architecture Group, email: Stephen.Childs @ cs.tcd.ie
Trinity College Dublin, Ireland web: http://www.cs.tcd.ie/Stephen.Childs
|