Newsgroups: comp.parallel.mpi
From: Richard Barrett <rbarrett@lanl.gov>
Subject: Re: Buffered or Blocking All-to-all
Organization: Los Alamos National Laboratory
Date: Mon, 08 Sep 1997 09:06:10 -0600
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Transfer-Encoding: 7bit
Message-ID: <34141462.217B@lanl.gov>


>  Is the MPI_ALLTOALL command blocking, or can it be
>buffered?

Implementation dependent and message size dependent. 
In most cases probably related to the behavior
of MPI_Send.

I believe the "Using MPI" book has an example of how to
implement a nonblocking broadcast, which you might look
at in order to write a nonblocking All-to-all.

See, even at 128 + 212 functions, 
MPI doesn't have _every_ possible function... :)

>I could probably acheive this
>effect by replacing the MPI_ALLTOALL call with a few "do"
>loops and many buffered Sends and Receives, but this latter
>method seems less clean and efficient.

Yeah, the word I've heard is that Bsend is a dog on most (all?)
platforms...

Richard