Re: Another proposal

Tony Skjellum (tony@Aurora.CS.MsState.Edu)
Mon, 3 Feb 1997 21:44:25 -0600 (CST)

I appreciate the concerns of Raja and Eric... I wish we could
do one thing though. Discuss models of the special collective hardware
they are thinking is realistic for next few years, so we can pattern
appropriate proposals. I wonder if communicators already defeat the
usefulness of certain classes, and I ask that the conversation be
elevated to models so we can speculate more usefully about changes
(if any) that support these up-to-now hypothetic technologies.

-Tony

On Mon, 3 Feb 1997, Tony Skjellum wrote:

> Date: Mon, 3 Feb 1997 17:50:07 -0600 (CST)
> From: Tony Skjellum <tony@Aurora.CS.MsState.Edu>
> To: Raja Daoud <raja@tbag.rsn.hp.com>
> Cc: Eric Salo <salo@mrjones.engr.sgi.com>, mpi-coll@mcs.anl.gov
> Subject: Re: Another proposal
>
> Of course, I wrote the text you are quoting. The purpose of this,
> implementation twins, is to guarantee that any implementation gives
> the application 1 synchronization, and N results. It is useful.
> -Tony
>
> On Mon, 3 Feb 1997, Raja Daoud wrote:
>
> > Date: Mon, 03 Feb 1997 14:54:54 CST
> > From: Raja Daoud <raja@tbag.rsn.hp.com>
> > To: Eric Salo <salo@mrjones.engr.sgi.com>
> > Cc: mpi-coll@mcs.anl.gov
> > Subject: Re: Another proposal
> >
> > >Seems like a relatively harmless addition to the standard, since we can just
> > >throw a loop around current MPI-1 calls
> >
> > :-)
> >
> > >shouldn't a high-quality application
> > >just manage a small pool of communicators for the best performance?
> >
> > Yup. In MPI-1, the story was that a high-qual impl. would pre-allocate
> > N context IDs at comm-create time, allowing the next N-1 comm-dup to be
> > lightweight (and reuse the context IDs, etc.).
> >
> > So a new function isn't needed. If the thinking is that this function
> > will help push implementors to do this kind of context ID mgmt, Eric
> > just showed how futile that can be.
> >
> > To make things really "fun" for users, implementors could make this new
> > function very heavyweight (N allreduce), while optmizing good old
> > MPI_Comm_dup() to pre-allocate context IDs. This would help push users
> > to stick to good MPI-1 functions. :-) :-) :-)
> >
> > --Raja
> >
> > -=-
> > Raja Daoud Hewlett-Packard Co.
> > raja@rsn.hp.com Convex Division
> >
>
> Anthony Skjellum, PhD, Associate Professor of Computer Science;
> Mississippi State University, Department of Computer Science & NSF ERC
> Butler, Rm 300, PO Box 9637, Corner of Perry&Barr, Mississippi State,MS 39762
> (601)325-8435 FAX: (601)325-8997; http://www.erc.msstate.edu/~tony;
> "Persistence is fertile." ; e-mail: tony@cs.msstate.edu; Try MPI!
>
>

Anthony Skjellum, PhD, Associate Professor of Computer Science;
Mississippi State University, Department of Computer Science & NSF ERC
Butler, Rm 300, PO Box 9637, Corner of Perry&Barr, Mississippi State,MS 39762
(601)325-8435 FAX: (601)325-8997; http://www.erc.msstate.edu/~tony;
"Persistence is fertile." ; e-mail: tony@cs.msstate.edu; Try MPI!