On Wed, 2011-12-14 at 16:16 -0500, Edsall, William (WJ) wrote:
> Unfortunately I had to clear them out to fit some jobs in.. but I have
> a similar comparison:
> 
> 
> ]# diagnose -p | more
> diagnosing job priority information (partition: ALL)
> 
> Job                    PRIORITY*   Cred(  QOS:Class)  Serv(QTime)
>              Weights   --------       1(    1:    1)     1(    1)
> 
> 258693                    10076     1.0(  0.0:100.0)  99.0(9975.)
> 258704                    10076     1.0(  0.0:100.0)  99.0(9975.)
> 258705                    10076     1.0(  0.0:100.0)  99.0(9975.)
> 258708                    10076     1.0(  0.0:100.0)  99.0(9975.)
> 258714                    10076     1.0(  0.0:100.0)  99.0(9975.)
> 258474                     9949     1.0(  0.0:100.0)  99.0(9849.)
> 258463                     9810     1.0(  0.0:100.0)  99.0(9709.)
> 
> These top jobs with 10,000+ are preempted, originally had 100
> priority. Currently they are suspended.

IIRC, the QUEUETIMEWEIGHT factor is multiplied by the number of minutes
the job has been eligible to run, and the default value for it is 1.

There are a couple different thing you could do here:

* Set QUEUETIMEWEIGHT to 0, which has the potentially unfortunate side
effect of not giving any priority to *any* jobs which have been sitting
for a long time.

* Set CREDWEIGHT to something larger than 1 (e.g. 10 or 100) to reduce
the relative effect of jobs' queue wait time relative to the priority
factor from the class.

* Use a QOS to adjust QUEUETIMEWEIGHT for the jobs in question.  (See
http://www.adaptivecomputing.com/resources/docs/maui/5.1.2priorityfactors.php#queuetimesubcomponent
 for details.)

        --Troy
-- 
Troy Baer, HPC System Administrator
National Institute for Computational Sciences, University of Tennessee
http://www.nics.tennessee.edu/
Phone:  865-241-4233


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

Reply via email to