[torqueusers] Queue Routing, Unexpected Behavior
Jason Williams
jasonw at jhu.edu
Wed Jun 24 12:06:15 MDT 2009
Jason Williams wrote:
> So I have set up some queue routing to test out an idea to have a
> different max wall clock time for a job depending on how many CPU's it
> requires. I figured, from reading the online docs, that this could be
> done via queue routing. I've put my queue config at the bottom of this
> email. The CPU count seems to be routing the job correctly, but the
> expected behavior of the different resources_max.walltime settings per
> queue routing destination seem to have no effect. Is what I am trying
> to do actually possible? One other setting that is set, but not
> included in the listing below for security reasons, is the acl_users
> setting. I don't see any reason why that would effect it, but I figured
> I would mention it for completeness. The job I tested requests 8 CPU's
> and 120:00:00 of wall clock. It winds up in the default_all queue which
> seemingly violates the resources_max.walltime setting. Is is a possible
> bug or am I not understanding how the routing works 100%...
>
> Thanks
>
After a bit of digging, it would seem the system WAS doing what it
should have been doing. The specifier in my example config for
'resources_max.ncpus' wasn't what I thought it was. 'ncpus' does NOT
equal any sort of derivative of '-l nodes5:ppn8' for example. I'm not
even sure there is a resource specifier that will catch that. If there
is, does anyone know what it is? I would like to make my routing limits
stringent enough to say 'jobs requesting 1 cpu' without using ncpus,
since specifying '-l nodes...' doesn't hit that rule since ncpus
defaults to 1. In order for ncpus to have an effect, you must request
them with '-l ncpus=32' for example. At least, that's the way I am
understanding this. Any suggestions/comments are welcome as I am pretty
well convinced that I may not be able to do what I need to do.
--
Jason Williams
Systems Administrator
Physics and Astronomy Department
Johns Hopkins University
More information about the torqueusers
mailing list