[Swift-devel] problem with coasters on pbs provider (on pads)

Mihael Hategan hategan at mcs.anl.gov
Thu Aug 5 16:43:08 CDT 2010


Oh, I see. PADS treats cores as nodes. Then what Mike says:
workersPerNode=1 and nodeGranularity=8.

On Thu, 2010-08-05 at 17:27 -0400, Glen Hocky wrote:
> with coasters some number of queue jobs spawn (16), and when 1 node
> (with 8 cpus) became available, the queuing system starts 7 coasters
> jobs which run 56 workers concurrently
> 
> 
> since this happens, what settings should I pick to have my desired
> behavior (i.e. 8 workers per node)
> 
> On Thu, Aug 5, 2010 at 5:23 PM, Mihael Hategan <hategan at mcs.anl.gov>
> wrote:
>         On Thu, 2010-08-05 at 17:14 -0400, Glen Hocky wrote:
>         > I'm having a problem running on PADS. It seems that when I
>         submit jobs
>         > with workerspernode=8, the queuing system doesn't pick up on
>         the fact
>         > that each job submitted by swift should have ppn=8
>         (specifically, that
>         > is missing from the qsub command registered by pbs)
>         
>         
>         It's not meant to.
>         
>         At some point in the past the meaning of "workerspernode" has
>         changed
>         from "start n instances of the worker" to "submit at most n
>         concurrent
>         jobs to one worker".
>         Since this applies to SMP "nodes" the end result is similar
>         (i.e. n jobs
>         per node), except only one worker.pl instance (and therefore
>         only one
>         TCP connection) is used per node.
>         
>         >
>         >
>         > when I do a qsub -f on my running jobs I get
>         >         submit_args = -A CI-CCR000013 -l
>         nodes=1,walltime=02:00:00,
>         >
>         >
>          size=1 /home/hockyg/.globus/scripts/PBS4482066898055181239.submit
>         >
>         >
>         > so there's no ppn=8 and i think it should also say size=8.
>         >
>         >
>         > the result is that I get 56 jobs running on one node
>         >
>         
>         
>         I lost you there. You get 56 j/n with coasters (which would be
>         bad) or
>         with the manual qsub (whose degree of badness I cannot
>         assess)?
>         
>         Mihael
>         
>         
> 
> 





More information about the Swift-devel mailing list