Newsgroups: comp.parallel.mpi
From: Steve Barnard <steve@megafauna.com>
Reply-To: steve@megafauna.com
Subject: Re: MPI question
Organization: megafauna
Date: Sat, 01 Jun 1996 19:53:32 -0800
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Transfer-Encoding: 7bit
Message-ID: <31B1103C.4295@megafauna.com>

Raja Daoud wrote:
> 
> Steve Barnard  <steve@megafauna.com> wrote:
> >blocking buffered sends (Bsend) were available, but now I see that a
> >non-blocking version (Ibsend) is there, too.
> 
> Note that Bsend doesn't block, and both Bsend and Ibsend consume
> resources (i.e. envelopes) and thus must abort/error if resources are
> depleted.  The suggested implementation of Bsend (page 36) calls Isend.
> The same issue of running out of resources is thus inherited.  Either
> the programmer is careful in the use of non-blocking calls (including
> "blocking" Bsend) or the application aborts/errors at some level as it
> is scaled.  This "some level" is implementation dependent.  So in this
> gray area, for non-scalable applications, it's just a matter of tuning
> parameters for particular systems.  Hope that helps.
> 
> --Raja
> 
> -=-
> Raja Daoud                                      Hewlett-Packard Co.
> raja@rsn.hp.com                                 Convex Division

Raja,

That is very interesting.  It's also very disturbing.  I would expect, 
at a minimum, a way to catch an error due to buffer resources being 
exhausted, and a way to handle it gracefully.

Anything less is a tacit admission that non-blocking sends are 
completely unreliable.

If this is true, and it sounds plausible, it's a serious problem with 
the current MPI standard.

	Steve Barnard

