Le 8 sept. 2005, à 12:05, Maarten Litmaath, CERN a écrit :
> Dear colleagues,
> please let us know what other kinds of job execution requirements for
> the
> batch system should be expressible in the JDL and subsequently
> forwarded
> to the batch system.
Start by forwarding everything which is already communicated to the
grid's WMS
(why not the JDL source?). Some batch systems might not be able to
handle all details, allow them to ignore parts without failing the job
where possible.
A comparison of what is supported by the batch systems used most often -
lsf, openpbs/pbspro, condor - might help to select reasonable candidates
for additional execution requirements.
As an example yet another batch system, here are some of the things
which our batch system can handle at the moment, many of them as
indicated
in the corresponding POSIX standard:
- Transfer of all or a subset of the user's environment variables as of
the moment when she enters the job submission command.
- Choice of the type of shell to execute with the script (bash, tcsh,
ksh etc.).
- Post mortem script
- LCG middleware version
- Platform information (Linux, RS/AIX, SUN/Solaris; AMD, Intel etc.)
- CPU execution time (not elapsed time)
- Memory limit (RAM)
- Scratch space limit
- stdio/stderr size limit
- "Dynamic virtual resources", example: SE available or not (this
is just a predefined character string used by the local
scheduler)
NOTE that this batch scheduler handles a heterogeneous machine park. No
subclusters. Those shown to the grid are merely virtual.
Cheers,
Rolf
|