[Mauiusers] Urgent jobs
t.wolf at tu-bs.de
Tue Apr 5 08:06:55 MDT 2005
Thanks for your suggestion! After having heard about those issues with
preemption, signals, moms and mpiexec I recalled your idea and finally
got you right. Now, I have two standing reservations: 8 nodes with
actually 2 cpus per node - np=4 in pbs' node file and 2 procs on every
node for the express queue (hiprio) and all of the other queues
(lowprio). Without the express queue the nodes get filled as before
(load per node equal 2). With additional urgent jobs a maximum of two
processes share one cpu (load per node equal 4).
I had the problem that the nodes were allocated within both reservations
regardless of the reservations of the other groups, so that there were
idle nodes and nodes running hiprio and lowprio jobs at the same time.
So I switched the NODEALLOCATIONPOLICY to CPULOAD.
Up to now, everything runs fine (for the last 20 minutes...)
More information about the mauiusers