[Swift-devel] more on # of coasters workers vs actual requested on ranger
Mihael Hategan
hategan at mcs.anl.gov
Tue Jul 21 13:20:32 CDT 2009
On Tue, 2009-07-21 at 13:13 -0500, Allan Espinosa wrote:
> aha.
>
> on Ranger the count clause , refers to the number of cpus hence when
> coasters is requesting for count=4 it only needs 1 node. if we want
> to do a workersPerNode=16 then we should manually specify host_count=4
> instead of count=4. or just use workersPerNode=1
Ah, right. I remember this funny problem.
Can you find out how well this is supported in general? The gram docs
are a bit vague:
(hostCount=value)
Only applies to clusters of SMP computers, such as newer IBM SP
systems. Defines the number of nodes ("pizza boxes") to
distribute the "count" processes across.
>
> i'll do more rsl exploration and probably play with the coaster's
> generation of GRAM2 requests.
>
> -Allan
>
> 2009/7/21 Mihael Hategan <hategan at mcs.anl.gov>:
> > On Tue, 2009-07-21 at 11:49 -0500, Allan Espinosa wrote:
> >> According to the gram logs, swift sends requests for blocks of 1, 2, 3
> >> and 4 nodes but SGE receives requests for four 1 node jobs. This
> >> maybe a GRAM2-SGE interaction problem. Is there a way to get the
> >> globus RSL files from swift so I can submit manually and verify this?
> >
> > In cog/modules/coaster/resources/log4.properties add:
> > log4j.logger.org.globus.cog.abstraction.impl.execution.gt2.JobSubmissionTaskHandler=DEBUG
> >
> > Then re-compile.
> >
> > But I don't think you need to go that far. Write your own RSL. In
> > particular I'd suggest trying with both jobType=multiple and without.
> >
> >
> >
> >
>
>
>
More information about the Swift-devel
mailing list