MPI-2 Process Management and mpiexec

[Pete Wyckoff]

trach-minh.tran at epfl.ch wrote on Mon, 19 Dec 2005 11:07 +0100:
> Thanks for your quick reply.
> 
> I don't think that I am capable of adding new features to mpiexec,
> my knowledge on TM things being practically zero!
> 
> For the moment, we would rather switch to mpich2 launcher
> for codes using MPI-2 dynamic process features .. while
> waiting that these features will be soon incorporated into mpiexec.
> 
> Thanks again for your great work!

In case you feel like playing around with it, I have checked into
the CVS basic support for MPI_Comm_spawn and the associated name
publishing and lookup routines you were asking about.  It works in
the sense that the most basic example I could find in the mpich2
source does successfully spawn, and non-spawning codes continue not
to fail.

So far I've been testing with mpich2-1.0.3 ch3:sock, but eventually
I'll get around to testing some different configurations.  Note that
OSU's mvapich2 (for InfiniBand) does not seem to support spawning.
Sometime I will spin a tarball release to ask for some wider
testing.

This work does not at all address the problem of how to get more
nodes from the resource manager.  You have to ask for all the
processors you will ever need in the PBS allocation, then use
mpiexec to spawn tasks only up to the number so allocated.  Going
beyond that will be a nice project for the Torque developers.  :)

		-- Pete