[torqueusers] tm enabled mpirun for mpt?
David.Singleton at anu.edu.au
Mon Sep 10 08:39:53 MDT 2007
It's not really possible to replace the MPT mpirun with a tm-enabled
one. There is a startup protocol between mpirun and the shepherd
MPI task that would be hard to replicate (you'll notice that there
are actually N+1 processes for an N process job - the extra one is
I think if you only have one host (and users do not specify the array
and host on the mpirun command line) then all the MPI processes are
children of mpirun and hence known to the MOM. So you should be
seeing all the job memory use. Unfortunately the memory use of MPT
processes is a bit strange since they map each others memory to do
zero-copy message passing. The virtual memory blows out with every
process having a VMA of 2-4GB for every other process (we currently
have a 200 cpu job where every process is supposedly 660GB). RSS is
also an overestimate since any page involved in a message pass will be
counted in the physical memory of at least two processes. So I would
imagine torque is actually overcounting job memory use, not
You could use MPI_MEMMAP_OFF to turn off memory mapping. Or you
could just run OpenMPI or LAM using shared memory segments although
you may lose some of the NUMA-awareness of MPT.
Brock Palen wrote:
> We have torque+maui on a SGI altix,
> We are enforcing memory use with maui and kill off jobs that go over the
> memory they requested. Problem is sgi's mpirun that came with MPT does
> not use tm therefor all the processes are outside the control of
> torque/maui resulting in users sucking up memory and over running users!
> Note we dont have cpu sets working its propack 5.
> Any insight from torque users would be wonderful.
> Brock Palen
> Center for Advanced Computing
> brockp at umich.edu
> torqueusers mailing list
> torqueusers at supercluster.org
More information about the torqueusers