[GE users] Scheduling policy question
prentice at ias.edu
Tue Sep 15 20:56:43 BST 2009
I have a 64-node cluster with 8 slots per node (dual quad-core
processors). Lately a lot of jobs are being submitted from
gridMathematica by several users. I have another user that wants to take
advantage of exclusive node use for his OpenMP programs. Unfortunately,
The default scheduler action (queue_sort_method = load)is creating a
problem for this scenario. Here's why.
gridMathematica breaks up it's jobs into separate, independent sub-jobs,
and submits them to SGE as individual jobs, so that if a user requests
32 processors in Mathematica, 32 separate jobs are submitted to SGE at
the same time.
Right now, I have one Mathematica user submitting jobs that request 2 or
3 processors, and then he submits another Mathematica job like that a
little later. The end result is that pretty soon, every node is running
Mathematica, but using only 2 or 3 slots. This prevents my OpenMP user
from running his job, because there are no nodes he can get exclusive
use of, even though only 1/4 - 3/8 of the slots in the cluster are being
The easiest way to fix this problem would be to set
queue_sort_method=seqno, but I'm afraid this isn't the best method,
either, since there probably will be times where I want jobs assigned to
the least-loaded node. I've read the sched_conf man page, but don't see
any obvious (simple) fixes.
If the Mathematica jobs were truly parallel, I could create a PE with
allocation_rule=$fill_up, which might help.
To unsubscribe from this discussion, e-mail: [users-unsubscribe at gridengine.sunsource.net].
More information about the gridengine-users