[GE users] "Mutually Exclusive" Queues

Brian R. Smith brs at usf.edu
Tue Sep 4 14:48:49 BST 2007


    [ The following text is in the "ISO-8859-1" character set. ]
    [ Your display is set for the "ISO-8859-10" character set.  ]
    [ Some special characters may be displayed incorrectly. ]

Hi all,

I'd like to be able to configure two queues, who have very similar 
hardware (the only difference is that the nodes reside on separate 
InfiniBand networks and hence cannot talk to each other -- one's an old 
SDR setup and the other is a newer DDR configuration) so that a user 
could request, say i_ib=true (which is common between both queues), but 
the job would not span across both queues.  Of course, an obvious 
solution is to have the user specify the queue, but there are many many 
problems with this.  I could also create separate complex values, i_ib1 
and i_ib2 to differentiate, but this would be additional worry for the 
users not to mention we would run into many of the same throughput 
problems posed by specifying a queue in a job script.  What I'd like to 
be able to do is have all slave tasks execute only in the first queue 
that satisfies the job requirement (this problem only really occurs for 
parallel jobs) e.g.

if queue1.slots_available >= job.requested_slots then
   run job & all job slaves in queue1
else if queue2.slots_available >= job.requested_slots then
   run job & all job slaves in queue2
else
   wait

where queues 1 and 2 both satisfy the job's requirements.  This sort of 
configuration should provide the best of both worlds by eliminating 
unnecessarily broad job constraints which cut jobs off from available 
hardware while ensure that slave tasks do not start up across different 
InfiniBand networks.  Is such a configuration possible?  How might I be 
able to achieve something like this?

Thanks,
Brian Smith

-- 
Brian R. Smith
HPC Systems Administrator
Research Computing, University of South Florida
4202 E. Fowler Ave. LIB618
Office Phone: +1 813 974-1467
Mobile Phone: +1 813 230-3441
Organization URL: http://rc.usf.edu

---------------------------------------------------------------------
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