Newsgroups: comp.parallel.mpi
From: harvey@epcc.ed.ac.uk (H Richardson)
Subject: Re: MPI_Ibsend()
Organization: Thinking Machines Corporation (c/o EPCC)
Date: Mon, 29 Jan 1996 14:28:41 GMT
Message-ID: <DLy5Jv.2wH.0.staffin.dcs.ed.ac.uk@dcs.ed.ac.uk>

In article <4e8ta0$qb9@imag.imag.fr>, christal@imag.fr (Michel Christaller) writes:
> 
> To my knowledge, one have to specify the size of the copy-buffer for
> all buffered sends, always. MPI doesn't manage any buffer in your place
> (it's one big difference from PVM).
> 

I don't think that MPI forces you to.  If you don't supply a buffer 
then MPI should operate as if you supplied a buffer of size 0.
However an MPI implementation is free to supply more buffering
(see sec 3.6) so your program may work.  Relying on such behaviour
would not be a good idea :-)

This brings me to another point.  Assume that I attach a buffer and 
then write to it at random places in my MPI application.  Is the
application legal?

Harvey

