Newsgroups: comp.parallel.mpi
From: raja@rsn.hp.com (Raja Daoud)
Subject: Re: MPI_RECV
Organization: Hewlett-Packard Co.
Date: 9 Jul 1998 23:45:13 -0500
Message-ID: <6o468p$n59@tbag.rsn.hp.com>

>but as an experiment.  I speculated that the "p4" implementation might
>be using shared memory for some kind of intermediate storage, and
>testing with MPI_Bsend might prove or disprove that hypothesis.

I see what you mean.  This assumes that MPI_Bsend would not use shared-
memory for intermediate storage until the receiver is ready, i.e.  all
MPI_Bsend messages are internally handled as sync-sends irrespective of
their lengths.  If that's not the case, and MPI_Bsend is internally
equivalent to an MPI_Isend from the attached buffer (as the example
implementation shows in the MPI standard), then it would still use up
shared-memory for short (eager) messages.

>What would a suitable "throttling policy" be?  I do understand your
>point about "high quality applications," but I don't see how one can
>throttle.

From the MPI side, it could detect when the destination buffer is full
(like the error message did) and hold back further messages until some
free shared-memory space is available.  From the application level, a
sliding-window protocol could be used between process-pairs (please
don't do that, keep your application as simple as possible :-) ).

Regards,

--Raja

-=-
Raja Daoud				Hewlett-Packard Co.
raja@rsn.hp.com				http://www.hp.com/go/mpi