Hi,
I configured maui to divide nodes in two pools using standing
reservations some time ago and, I admit, It works. My definitions look
this way:
SRCFG[test] PERIOD=INFINITY HOSTLIST=n96,n97 TIMELIMIT=15:00*
ACCESS=DEDICATED QOSLIST=test
SRCFG[batch] PERIOD=INFINITY HOSTLIST=n[1-9]$,n[1-9][012345]$
ACCESS=DEDICATED QOSLIST=light,bulk,test
SRCFG[v40z] PERIOD=INFINITY HOSTLIST=n10[123456789]$,n11[012]$
ACCESS=DEDICATED QOSLIST=v40z,test
So I have three not overlapping reservations. Every job which fits into
one of this reservations wont't fit in any other of them. It works. But
scheduling, ehm, works quite illogical.
In short - only the highest priority job is ever started, jobs with
lesser priority waiting in the queue won't start even when there are
free resources in the reservation they fit in and the highest priority
job, waiting to be started, won't. Simply from logical point of view I
have two pools of nodes, but I can't force maui to stop treating them as
one big pool.
Example:
54810 user2 Idle 16 2:00:00:00 Mon Apr 3 10:56:04
54793 user2 Idle 8 12:00:00 Mon Apr 3 10:57:13
Job 54810 has higher priority than 54793 and should be started first.
Job 54810 fits into batch reservation, job 54793 fits into v40z
reservation. There are enough free resources in v40z reservation to
start job 54793. Job 54810 has to wait for free resources. Conclusion -
job 54793 can, and should, be started at once. The way it works - lesser
priority job will never start before higher priority one. I can see
shown situation every several hours.
Anyone knows how to overcome this issue? It's quite fustrating to
correct scheduler by hand every several hours...
There are two more 'issues' (well, bugs). First - when maui is started
it sets reservations on online nodes only. It means that if node is
down, maui is started and node is up later, it will have na reservations
at all. Second - CLASSCFG parameter in SRCFG declaration doesn't work -
being in class allowed to use reservation makes maui to return that job
is not allowed to use that reservation while in fact it should allow it.
Marcin Mogielnicki
_______________________________________________
mauiusers mailing list
[email protected]
http://www.supercluster.org/mailman/listinfo/mauiusers
