Hi,
At Cray, we have a a Cluster Compatibility Mode (CCM)
which runs under PBSPro. If we do a qsub,
we get an allocation of compute nodes and the first
node in the list of allocated nodes is considered the
"head node" of the "Cluster".
The list of allocated nodes is contained in
the file $PBS_NODEFILE.
To launch a command on the head node we type:
ccmrun COMMAND
In the normal PBSPro environment, we launch
the OpenMPI mpirun as:
ccmrun ${OpenMPI_Install}/bin/mpirun <options ..>
-hostfile host.list -np $NCPUS ./hello_mpi
This all works great with the "mpirun" command
launching on the "head node" and the the mpirun
command using ssh/rsh to launch other processes
on the other nodes in the virtual cluster. Everything
works.
On Cray sites, that use PBSPro/Torque/Moab, OpenMPI
detects that Torque/Moab is present and attempts
to use Torque/Moab to launch on other nodes. Because of
the way Cray has Torque/moab configured, this doesn't work.
This is not a problem with OpenMPI.
Is there an option to OpenMPI to tell it to ignore
Torque/Moab and use rsh/ssh and the -hostfile/hosts
even if Torque/Moab is detected?? I've looked through
the README and the FAQ but I haven't seen anything.
Many thanks,
Dave
--
CCCCCCCCCCCCCCCCCCCCCCFFFFFFFFFFFFFFFFFFFFFFFFFDDDDDDDDDDDDDDDDDDDDD
David Whitaker, Ph.D. whitaker_at_[hidden]
Aerospace CFD Specialist phone: (651)605-9078
ISV Applications/Cray Inc fax: (651)605-9001
CCCCCCCCCCCCCCCCCCCCCCFFFFFFFFFFFFFFFFFFFFFFFFFDDDDDDDDDDDDDDDDDDDDD
|