[Mauiusers] array jobs and priorities

Steve Crusan scrusan at ur.rochester.edu
Mon Aug 8 14:36:22 MDT 2011

Hash: SHA1

On Aug 8, 2011, at 4:23 PM, Andrus, Brian Contractor wrote:

> All,
> I am seeing an odd thing in array job prioritization.
> It appears that when someone submits an array job with a limit on concurrent runs (eg: -t 1-1000%25), the jobs that are queued continue to accrue priority, but do not run because the limit is stopping them.
> This is problematic when there are other jobs in the queue. I end up with MANY array jobs eligible to run, but do not because of the concurrency limit. Other jobs (stuck with a lower priority) do not run because they are waiting for the higher priority array jobs to run…

I'm not actually 100% sure about this, but  you can set a limit on the number idle jobs, meaning if a user submits 1000 jobs, only X number of jobs accrue priority, and the rest are blocked. I don't think blocked jobs accrue priority. If I check my queues with showq -i, idle jobs accumulate priority, but when I use showq -b, I don't see any priority listings for blocked jobs. 


If I'm not mistake about idle jobs vs. blocked jobs in terms of priority, you can use the MAXIJOB=X parameter to set this.

> Is there a way to have array jobs accrue priority at a lower rate than standard jobs? Or if a job is deferred or otherwise held off due to an array throttle, it does not accrue priority?
> Finally, is there a variable for interactive jobs I can use for setting priority? I wish to have interactive jobs get the highest priority.

We map any jobs submitted to our interactive queue to a QOS, which then gives each job a starting priority. I'm starting to move towards a QOS approach with almost all of my queues, whereas TORQUE controls access to the queue via acl groups & hard limits on job geometry, and then the scheduler handles everything else through QOS.

> Brian Andrus
> ITACS/Research Computing
> Naval Postgraduate School
> Monterey, California
> voice: 831-656-6238
> _______________________________________________
> mauiusers mailing list
> mauiusers at supercluster.org
> http://www.supercluster.org/mailman/listinfo/mauiusers

 Steve Crusan
 System Administrator
 Center for Research Computing
 University of Rochester

Version: GnuPG/MacGPG2 v2.0.17 (Darwin)
Comment: GPGTools - http://gpgtools.org


More information about the mauiusers mailing list