Re: [ESPResSo] Parallel Simulations and System Initialization

[Olaf Lenz]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: RIPEMD160

Hi Lorenzo!

Lorenzo Isella wrote:
| I am asking really two different things.
| (1) I am completely new to running codes in parallel. My machine runs
| Debian Testing and is a  64-bit machine with two CPU's.
| I installed the OpenMpi library and configured espresso with the
| command (without any special options, I thought the script may pick up
| the MPI installation by itself):

Indeed, OpenMPI is not yet directly supported by ESPResSo. This means,
that it is not automatically recognized and can't set all the necessary
environment variables and compiler flags. We should definitely add this
to ESPResSo.

Still, it is no problem to use ESPResSo with OpenMPI - many people do
so, including myself.

Getting ESPResSo to run with MPI requires two things: an MPI compiler,
and the MPI runtime environment. The MPI compiler "mpicc" was
automatically recognized in your setup, and from your mailing I conclude
that it worked fine (all that happened is that instead of "gcc", the
compiler "mpicc" was used).

Running ESPResSo in MPI is a bit more complicated. Usually, MPI has to
start some kind of runtime environment, only then you can actually run
an MPI program. I do not know exactly how this works in OpenMPI - the
clusters that I work on have the MPI runtime environment set up and
running by the administrators.

Usually, you first have to start the runtime environment (mpiboot or
something), and afterwards you have to start your program using
"mpirun". Note, however, that ESPResSo tries to do that last step for
you. "Espresso" is a shell script that calls "mpirun" and starts
"Espresso_bin".

It might be that you will have to consult the OpenMPI docs, if not
anyone else from the list answers your question. Try "man mpirun" first.

| $ ./Espresso test_interaction.tcl
| libibverbs: Fatal: couldn't read uverbs ABI version.
[clip]

I would guess that these other errors emerge from the missing MPI
runtime environment. So try first to get OpenMPI running correctly, then
try again.

As for your second question, I don't know exactly. Probably someone else
can answer the question. I would guess that you have to give a seed for
each processor, as they have independent random number generators.


Cheers
        Olaf
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.6 (GNU/Linux)
Comment: Using GnuPG with Mozilla - http://enigmail.mozdev.org

iD8DBQFHpi1ntQ3riQ3oo/oRA22FAKCz6CADkWRx8l9h+9I35S5DvVr/uQCdGNIH
hfrM9tXpkbsGpoHKefMEmVQ=
=PYGY
-----END PGP SIGNATURE-----