equivalent of mpirun -nolocal option?

[Ron Brightwell]

> > Is there an mpiexec equivalent to the mpirun -nolocal option?
> > I'd like to be able to start a job where the mpiexec process
> > is on a separate host from the MPI application process(es).
> 
> Ah, no.  It's not even possible to hack via config-file stuff
> for a generic MPI code either, unless you're allowed to recompile
> the code to exit task 0, form a new smaller communicator, and continue.
> 
> But you could toss in this option toward the end of get_hosts().
> Just chop off tasks[0] while the name matches the original tasks[0].name
> and move up everybody else in the array, reducing numtasks.  I'll
> fold it in to the distro if you get it to work (mostly) and find it useful.
> 

Ok. Thanks.

Our instrumentation of the MPICH/GM unexpected message queue consistently
shows strange data for rank 0.  For most of the codes that we've run, the
maximum length of the unexpected queue and the average traversal length of
the unexpected queue are an order of magnitude greater for rank 0.  We were
thinking that the competition for cycles with mpiexec might be a possible
cause, so we were looking for a quick way to check that out.

-Ron