mpiexec - PMI job lauch (mpich2)

[Pete Wyckoff]

johnnydevaprasad at gmail.com wrote on Thu, 13 Sep 2007 12:29 +0200:
> When i try to run an mpich2 job with the following line
>    mpiexec --comm=pmi -n 2  ./test  (because my default mpiexec is built
> with --with-default-comm=mpich-p4)
>       (I have also tried the above after rebuilding mpiexec with
> --with-default-comm=mpi)
> 
> mpiexec seems to spawn two different processes which dont belong to the same
> process group. The output of my program is
>    Hello world from process 0 of 1 from node002
>    Hello world from process 0 of 1 from node001

Sorry I missed this 10 days ago.  Mailman sits on incoming messages
when the sender isn't a list member and I usually get around to
approving them manually.  The presence of HTML hid the approval
message deep in my spam folder.

johnnydevaprasad at gmail.com wrote on Thu, 13 Sep 2007 14:37 +0200:
> I just missed out on some more information in my previous email.
> 
> I am using the mpich2 that is built with smpd as the process manager. Does
> mpiexec (yours) support this.
> Also my program works fine when i use mpich2 built with the mpd process
> manager. So it appears that the current version of mpiexec does not support
> the smpd process manager.
> Any feedback would be greatly appreciated.

This is the key insight.  I'm not sure if anyone has before tested
mpiexec with mpich2's smpd, or what the advantages of using smpd
are.  All I get from the docs is:

    SMPD is the primary process manager for MPICH2 on Windows.

You could run "mpiexec -v -v -v" to watch the PMI negotiation with
your SMPD version, then read the source in mpich2 to get more
information and write a patch to mpiexec to accommodate it.  Or you
could just decide not to use SMPD ever again.  You can guess what my
vote would be.  :)

I'll add a FAQ entry now that you've discovered this for us.

		-- Pete