On Apr 22, 2008, at 8:48 AM, Hans Wurst wrote:

Make sure you're using the "right" mpirun. If you have both MPICH and
Open MPI installed, you need to install them into the different
directories.

I am sure. I have separate VMware-instances installed in order to avoid any interdependencies. Although I'd like to have MPICH and OpenMPI installed on the same machines in the future to keep my efforts for maintenance small, I invested these endeavors for the first.

I would be grateful, if you could advice me about a clean setup for a machine with both MPICH and OpenMPI installed, but I don't want to mix up things here.

All you need to do is install them both into separate prefixes (e.g., / opt/openmpi and /opt/mpich). Then when you're using them, ensure that your PATH / LD_LIBRARY_PATH is set consistently across all nodes, or if using Open MPI, use the options listed in the "What if I can't modify my PATH and/or LD_LIBRARY_PATH?" FAQ entry (I'm particularly fond of the --enable-mpirun-prefix-by-default option).

As for not finding the "orted" on the remote nodes, it simply means
that your PATH / LD_LIBRARY_PATH is not set properly on the remote
nodes.  See:

http://www.open-mpi.org/faq/?category=running#run-prereqs
http://www.open-mpi.org/faq/?category=running#adding-ompi-to-path
http://www.open-mpi.org/faq/?category=running#mpirun-prefix

My PATH should be set properly. As I wrote before:

"
The "orted"-executable resides as well as the "mpirun"- and "mpiexec"-executables in the directory /home/stephan/openmpi- install. "orted" is also found by "which orted".
"

I've exported it with the following command:

 export PATH=/home/stephan/openmpi-install/bin:$PATH

Do I get you right:

You advice me, to export the /home/stephan/openmpi-install/lib- directory to LD_LIBRARY_PATH too? I can try this at home this evening.

Yes. But the point is that if you're not using something like -- enable-mpirun-prefix-by-default, then you need to set this PATH and LD_LIBRARY_PATH on *all* nodes (e.g., in your shell setup files). It is not sufficient to export PATH on just the node where you invoke mpirun. See the FAQ entries for more details.

- "--hostfile FILENAME", where FILENAME is a list of a hostfile as
described in the FAQ.

    mpirun --hostfile my_hostfile.txt ...

Yes, my hostfile is named mpd.hosts and has the content

NodeA
NodeB

, but I guess, that should fit the requirements, right?

Yes, that should be fine. Hosts listed without slot counts are assumed to have a slot count of 1.

--
Jeff Squyres
Cisco Systems

Reply via email to