Itay M wrote:
We've tried the new configuration (unset resources_default.ncpus and
unset resources_max.ncpus; from from queues and server levels as well)
in the last few days and here are the results:
I suppose you did check with qstat -f that 'ncpus' is not mentioned
anywhere any longer?
* For the first time we were able to see that jobs are backfilled! It
never happend before, and this is a major improvment. Though we saw it
only in one of our queues (named 'b_que') it might have happend in other
queues as well (we couln'd verify it yet).
* But - the 'insufficient idle procs available' problem is still there.
For example, at the moment showq shows that there are plenty of non-busy
processors ('65 of 84 Processors Active'), but checkjob says for
queued jobs that:
checking job 228665
State: Idle
Creds: user:b group:b class:b_que qos:hi
WallTime: 00:00:00 of 00:05:00
SubmitTime: Tue Jan 29 19:47:04
(Time Queued Total: 00:07:49 Eligible: 00:07:16)
Total Tasks: 1
Req[0] TaskCount: 1 Partition: ALL
Network: [NONE] Memory >= 0 Disk >= 0 Swap >= 0
Opsys: [NONE] Arch: [NONE] Features: [NONE]
Dedicated Resources Per Task: PROCS: 1 MEM: 512M
IWD: [NONE] Executable: [NONE]
Bypass: 0 StartCount: 0
PartitionMask: [ALL]
Flags: RESTARTABLE
PE: 1.00 StartPriority: 1007
job cannot run in partition DEFAULT (idle procs do not meet requirements
: 0 of 1 procs found)
idle procs: 12 feasible procs: 0
>
> :(
> What should I check next?
Maybe it has something to do with the MEM requirement (just a wild
guess... but try removing it). What does diagnose -n say for a node
which is incorrectly rejecting the job? Does it have enough free
"tokens" (not sure if this is what they are called officially) to run
the job in this b_que class?
Regards,
Jan Ploski
_______________________________________________
mauiusers mailing list
mauiusers@supercluster.org
http://www.supercluster.org/mailman/listinfo/mauiusers