looking into this more

[Pete Wyckoff]

dskinner at nersc.gov wrote on Fri, 22 Jul 2005 11:59 -0700:
> "The permanent solution is to use this patch to mvapich 0.9.5 then to
> fix mpiexec to recognize the new protocol version number. No one is
> currently working on this task, however."
> 
> It looks like the patch is already in 0.9.5-117. Can I volunteer to work
> on getting mpiexec to work with version 3?

The patch is in mvapich?  Great.  I talked to the mvapich guys a bit
back when they released 0.9.5, complaining that they changed the
protocol without changing the version number, and introduced some
changes that would hurt scalability in the process.  They did come
up with a patch from my suggestions that looked good (the one
referred to above), but I never took the time to test it---we still
run 0.9.4 here.

To start, you might compare pmgr_client_mpirun_rsh.c from an 0.9.4
and a late 0.9.5 to see if you can pick out the differences, then
work them into mpiexec's ib.c as a new version three.  I think they
added a few fields to ship pid lists around to support certain SMP
cases.

You might take a look at the patch Anton put together back in April.
It never got into mpiexec because of this problem of detecting which
mvapich we were talking to.  I didn't want to release something
where the user had to compile specially according to what version of
mvapich they happened to be using; that's what protocol version
numbers are all about (ipv4 vs ipv6, e.g.).  Note he's got two
concerns: the MPIRUN_MPD and MPIRUN_PROCESSES env vars (hopefully
neither is really required, we're not mpd and the other will scale
terribly), and the HAVE_IB_PIDS extra communication section in ib.c
that will get turned into (version == 3) blocks.

See this thread:
    http://email.osc.edu/pipermail/mpiexec/2005/000460.html
and this patch in particular:
    http://email.osc.edu/pipermail/mpiexec/attachments/20050425/36fc4d08/mvapich_pids.bin

I do look forward to merging in a nice clean patch to support this
latest mvapich.  Thanks for offering to help!

		-- Pete