[GE users] scheduler chooses not the best node in pe smp

reuti reuti at staff.uni-marburg.de
Thu Jul 8 16:17:40 BST 2010


Am 08.07.2010 um 11:49 schrieb jochen:

> Reuti, your proposal sounds good but the idea behind my question is the following.
> He have a very heterogeneous environment. He have now some nodes with 2, some with 4
> and some nodes with 8 cores. If y user request 4 cores, the job should use one node with
> 4 cores or (if none is available) 2 nodes with 2 cores.
> We do not want to waste 8 core power with a 4 core job running on a 8 core node.

you could run 2 jobs with 4 cores on such a machine then.


Anyway, what you can do, is to have multiple PEs with a fixed allocation rule:

8 core nodes: PE: mpi8 = allocation_rule 8
4 core nodes: PE: mpi4 = allocation_rule 4
2 core nodes: PE: mpi2 = allocation_rule 2

In the queue configuration you can use hostgroups to bind the different PEs to different types of machines:

pe_list NONE,[@core8=mpi8],[@core4=mpi4],[@core2=mpi2]

Then you can submit with:

$ qsub -pe "mpi*" 4 test.sh

SGE will discover, that this job can't be scheduled to the 8 core nodes, leaving you with mpi4 and mpi2. For this to put in an order, you can define different sequence numbers for the hostgroups in the queue definition (with a similar syntax like outlined above) and change the scheduler sort order to "seqno", i.e. it should first try the larger nodes.

-- Reuti

PS: please quote always the mail you reference to (there is a button for it in the web interface). It's really hard for me to have all threads in mind or to lookup the progress of discussion in the mail archive.

> Regards, Jochen
> ------------------------------------------------------
> http://gridengine.sunsource.net/ds/viewMessage.do?dsForumId=38&dsMessageId=266673
> To unsubscribe from this discussion, e-mail: [users-unsubscribe at gridengine.sunsource.net].


To unsubscribe from this discussion, e-mail: [users-unsubscribe at gridengine.sunsource.net].

More information about the gridengine-users mailing list