Torque 2.3.0 cpusets broken [PATCH] (was Fwd: [torquedev] Quick
test report of cpuset support)
David.Singleton at anu.edu.au
Mon Apr 7 17:02:54 MDT 2008
Yeah, I would suggest the right place for the subcpuset creation
is in the application layer - best if MOM just creates the
container cpuset. Consider the case of a hybrid OpenMP/MPI app
wanting two 4-thread tasks on each of your 8-core nodes. That's
a subcpuset configuration that might only be determinable by
the application at runtime - couple climate codes can have
nprocs/nthreads configuration in input files.
Chris Samuel wrote:
> ----- "Steve Snelgrove" <ssnelgrove at clusterresources.com> wrote:
>> I have applied the supplied patch and built a new snapshot with this
> Thanks Steve.
> It has exposed an oddity in OpenMPI though, whereas
> both the OSC mpiexec and pbsdsh do TM requests for
> each vnode separately OpenMPI appears to make one TM
> request *per node*, and then fork the tasks on that node.
> This means that with cpusets working all the OpenMPI
> processes for a job on a node end up locked to the
> same core! :-(
More information about the torquedev