Hi All.

We have a test bed setup running Maui 3.3.1 running with Torque 2.5.9.

We are testing PREEMPTOR and PREEMTEE setup with Maui and all works well until Maui starts new backfill jobs.

The new backfill jobs started by Maui will then not let PREEMTOR jobs run.    Is this a bug or are we missing a priority?

Our Torque nodes file have:

compute-1-1 np=64 chem free
compute-1-2 np=64 chem free
compute-1-3 np=64 chem free
compute-1-4 np=64 chem free
compute-1-5 np=64 chem free
compute-1-6 np=64 chem free
compute-1-7 np=64 tw free
compute-1-8 np=64 tw free
compute-1-9 np=64 tw free
compute-1-10 np=64 tw free

Torque "default" queue is "free":
set server default_queue = free
set server resources_default.nodes = 1


We fill up the queue with 10 63-core jobs and 15 1-core jobs as "PREEMPTEE" jobs "free" queue.   Maui starts 10 63-core jobs and 10 1-core jobs (leaving 5 1-core jobs in the queue).

As user "tw" I run:
            qsub -I  -q tw  -l nodes=4:ppn=64     ( PREEMPTOR job )
it PREEMPTS correctly.   Everything works as expected.

However, if I run:
            qsub -I  -q tw  -l nodes=4:ppn=62     ( PREEMPTOR job )

It PREEMPTS correctly and Maui then starts 4 new backfill 1-core jobs (since there are now 4 new 1-core CPU idle).

But here is the problem.    If I then re-run the command:
          qsub -I   -q tw  -l nodes=4:ppn=64      
( PREEMPTOR job)

The job will NOT run.   The 4 new 1-core jobs Maui started are not SUSPENDED by Maui.

Is this a bug?


------------------- maui.cfg file ------------------------------------
NODEALLOCATIONPOLICY  MINRESOURCE

PREEMPTPOLICY  SUSPEND
ENABLEMULTIREQJOBS  TRUE

# QOS
QOSWEIGHT      10000

QOSCFG[high] PRIORITY=1000  QFLAGS=PREEMPTOR QTWEIGHT=100
QOSCFG[low]  PRIORITY=0      QFLAGS=PREEMPTEE QTWEIGHT=0

#The default queue is called "free"
CLASSCFG[free] QDEF=low
CLASSCFG[free] DEFAULT.FEATURES=free

# TW Queue
SRCFG[tw]    PERIOD=INFINITY
CLASSCFG[tw] QDEF=high
CLASSCFG[tw] DEFAULT.FEATURES=tw





_______________________________________________
mauiusers mailing list
[email protected]
http://www.supercluster.org/mailman/listinfo/mauiusers

Reply via email to