[torqueusers] queue resources_min.nodect ignored from routing
garrick at clusterresources.com
Mon Nov 6 09:25:26 MST 2006
On Mon, Nov 06, 2006 at 10:14:01AM -0500, Bill Wichser alleged:
> pbs_version = 2.0.0p8
> linux w/2.6 kernel
> I have 3 queues set up: default, small, and large.
> Default is a routing queue which just sends to small then large.
> small is set up
> set queue small resources_max.nodect = 16
> to allow jobs smaller than 16 nodes to run there. It is further set in
> maui to allow no more than 112 nodes to be active. This functions fine.
> large is set up
> set queue large resources_max.nodect = 128
> set queue large resources_min.nodect = 32
> to require 32 node jobs and better, realizing that a gap exists between
> 17 and 31 node requests. This works fine as well.
> What doesn't work is when the small queue is full, ie the maui limit has
> been reached, and single processor jobs move into the large queue until
> that too is full. I expected that Torque would stop this as the job has
> only requested a single node and therefore would not meet the
> resources_min.nodect=32 limitation, but jobs just continue to fall into
> this queue until this fills.
> One note is that these job scripts do NOT specify how many nodes are
> being requested. They only specify a wallclocktime. But in the Torque
> server itself I do specify a line:
> set server resources_default.nodect = 1
> which should be taking care of this situation.
> Am I simply missing something else or is there perhaps something wrong
> with this version of the scheduler?
nodect is ment to be calcuated from nodes, so you wouldn't set a default
nodect. Set a default nodes.
Also, maui configs don't influence job routing in TORQUE.
More information about the torqueusers