[GE users] Preemption

Brian R. Smith brian at cypher.acomp.usf.edu
Fri Oct 29 18:46:40 BST 2004


Hey,

I've asked this question once and I'm going to try to ask it again, this
time phrased a little more intelligently (hopefully).

We've a cluster with 12 nodes and a head node.  Currently, I have SGE
6.0u1 configured to run as the job scheduler.  My parallel environment
for MPICH is configured correctly.  Jobs run perfectly and I have no
problems.  However, I need to create a scheduling policy that will
satisfy the demands of two different research groups.

I have two problems:

1) Regardless of the queue priority and the queue used, multi-processor
jobs ALWAYS have higher priority and are run first.  I have a 'group1.q'
and an 'all.q' that is subordinate to 'group1.q'.  If I submit a
parallel job to 'all.q', it MUST have lower priority than any job that
gets submitted to group1.q.  How can this be accomplished?

2) My two research groups would like to divide the cluster resources in
half, e.g. one group gets 6 nodes and the other group gets the other 6
nodes.  However, they also want the option to run on ALL of the nodes.
The caveat is that if one group's job prevents the other group from
running jobs, the scheduler should remove enough offending jobs from the
queue to allow the rightful group to run their job(s).

Currently, I have three queues, group1.q, group2.q and all.q.  all.q is
subordinate to both group1.q and group2.q via the queue_conf parameter
subordinate_list.  I would like my group1.q and group2.q queues to
delete any job running in all.q that is preventing a scheduled job to
the two high priority queues from running.  How can this be
accomplished?

People have given me the info on subordinate queues.  I had a suggestion
with regards to job relocation (which I found is not necessary as I only
have to delete the job).  Most of this information has _appeared_ sound
but upon implementation (due to incorrectness/incompleteness or my own
naivete) never seems to work.  

I have been working on this deployment for a couple of months now and
could really use some assistance.  If you have any configuration
information, how-to's, documentation, or anything that is remotely along
these lines and would be willing to share it, I would really appreciate
it.  


Brian Smith


---------------------------------------------------------------------
To unsubscribe, e-mail: users-unsubscribe at gridengine.sunsource.net
For additional commands, e-mail: users-help at gridengine.sunsource.net




More information about the gridengine-users mailing list