M. Stavrianakou, ATLAS software workshop, 97/09/04

LSF status update

LSF endorsed by FOCUS

to be adopted as the single batch scheduler at CERN,

with full migration of the LHC experiments by end 1997

SHIFT31 and SHIFT32 now running LSF
but CHANGES and/or PROBLEMS...

LSF authentication: klog command now creates both AFS token and kerberos ticket
WORKDIR: cd $WORKDIR when the directory is not explicitly specified by the user does not work -> solved?
LSF time limits i.e. NCUs vs native seconds: requests for clarifications...
Logfile viewing for jobs running on SHIFT: still no fix?
Job scheduling algorithm and latency on SHIFT31 and SHIFT32: complaints esp. for short (8nm) jobs: even when short queues are empty, jobs will not start if all available slots are occupied -> the mechanism on shift machines (which at present is NQS-flavoured) needs to be improved: execution should be based on load and queue priority rather than on available job slots (need to count number of jobs that are staging - not yet possible).
Killing jobs: severe problems reported by Dario

Conclusion:

LSF has been `misbehaving'... The various problems need to be solved urgently - certainly before any further migration! However, we must not forget that it is a very powerful system of which we've had very positive experience (at least until August!)