Am 07.06.2008 um 00:06 schrieb SLIM H.A.:
The btl interface is used. The nodes in both subclusters have the
same architecture so no need for two binaries. The MPI libraries
for the interconnect can be selected at runtime, for example
through the value of the OMPI_MCA_btl environment variable. However
that variable should then somehow be set once SGE has decided on
which subcluster to run the job, thus avoiding it to be set in
advance when submitting. I tried a couple of things like different
PEs but there does not seem to be an elegant solution; I mean
something completely transparent to the user like
A transparent solution would be to create a "starter_method" in the
parallel queue, which will check the node that it's running on and
hence you know which type of interface to use. Export the variable
and exec the real script; i.e. something like:
#!/bin/sh
if [ "${HOSTNAME%%[0-9]*}" = "node_ib" ]; then
export OMPI_MCA_btl=tcp
else
export OMPI_MCA_btl=openib
fi
exec "$@"
I think you get the idea of the script. Maybe you have some special
names for these machines. Another indicator would be the $PE or
$QUEUE of SGE. If you want to be sure to get only nodes of one and
the same type (but you don't care which at submission time), you will
need to setup something like:
http://gridengine.info/articles/2006/02/14/grouping-jobs-to-nodes-via-
wildcard-pes
anyway.
HTH - Reuti
mpirun -np $NSLOTS executable
as I have seen in some of the OpenMPI FAQs.
Thanks
Henk
From: users-boun...@open-mpi.org on behalf of Reuti
Sent: Fri 6/6/2008 6:47 PM
To: Open MPI Users
Subject: Re: [OMPI users] using OpenMPI + SGE in a heterogeneous
network
Am 06.06.2008 um 19:31 schrieb Patrick Geoffray:
> SLIM H.A. wrote:
>> I would be grateful for any advice
>
> Just to check, you are not using the MTL for MX, right ? Only the
BTL
> interface allow to choose between several devices at run time.
At least there would be the option to built two binaries and decide
at runtime which to use in your jobscript - depending on the machine
you landed on.
-- Reuti
_______________________________________________
users mailing list
us...@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/users
_______________________________________________
users mailing list
us...@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/users
