[mpich-discuss] Oversubscription with Nemesis

[Jed Brown]

The problem is that the slowdown is orders of magnitude, making it truly
unusable. That is very different from not useful for performance analysis
and not providing a speedup from smaller numbers of processes.

PETSc's configure has similar options, but the default is sock so very few
people configure with nemesis.
On Sep 11, 2012 4:50 PM, "Jeff Hammond" <jhammond at alcf.anl.gov> wrote:

> Hi Jed,
>
> I don't see the problem.  You assert that people who oversubscribe
> don't care about performance, and all that MPICH2 is doing wrong with
> Nemesis right now has to do with performance, right?
>
> Anyone who cares about performance shouldn't be relying upon PETSc to
> build MPICH2, so I see no problem in PETSc continuing to build MPICH2
> in a conservative way that supports oversubscription effectively.
>
> Does the PETSc build system support options e.g.
> --download-mpich=nemesis vs. --download-mpich=sockets?  That would
> seem to resolve all of your issues with minimal effort.
>
> Best,
>
> Jeff
>
> On Tue, Sep 11, 2012 at 2:11 PM, Jed Brown <jedbrown at mcs.anl.gov> wrote:
> > I know this has been filed as a ticket, but it seems to be postponed
> > quasi-indefinitely at this point.
> >
> > http://trac.mcs.anl.gov/projects/mpich2/ticket/79
> >
> > I'd like to emphasize that over-subscription is extremely commonly used
> > during initial development and method experimentation (not performance
> > studies, obviously) and consequently, PETSc continues to use ch3:sock as
> the
> > default for new installs (via --download-mpich, which is used by a large
> > fraction of PETSc users). It appears that you are trying to wean the
> > community off of sock in favor of nemesis, but as long as this bug is not
> > fixed, you will continue to get thousands of new sock installs.
> >
> > _______________________________________________
> > mpich-discuss mailing list     mpich-discuss at mcs.anl.gov
> > To manage subscription options or unsubscribe:
> > https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss
> >
>
>
>
> --
> Jeff Hammond
> Argonne Leadership Computing Facility
> University of Chicago Computation Institute
> jhammond at alcf.anl.gov / (630) 252-5381
> http://www.linkedin.com/in/jeffhammond
> https://wiki.alcf.anl.gov/parts/index.php/User:Jhammond
> _______________________________________________
> mpich-discuss mailing list     mpich-discuss at mcs.anl.gov
> To manage subscription options or unsubscribe:
> https://lists.mcs.anl.gov/mailman/listinfo/mpich-discuss
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/mpich-discuss/attachments/20120911/4cb2058a/attachment.html>