X-Mailer: QUALCOMM Windows Eudora Pro Version 4.2.2 
Date: Thu, 10 May 2001 13:21:38 -0500
To: mpi-21@XXXXXXXXXXXXX
From: William Gropp <gropp@XXXXXXXXXXX>
Subject: Re: MPI_FINALIZE
Cc: <mpi-21@XXXXXXXXXXXXX>,"Jeff Squyres" <jsquyres@XXXXXXXXXXX>,In-Reply-To: <001e01c0d97c$12fa9cc0$6163a8c0@XXXXX>
References: <200105100849.f4A8npK08386@XXXXXXXXXXXXXXXXXXXXXXXXXX>
Mime-Version: 1.0
Content-Type: text/plain; charset="us-ascii"; format=flowed
Sender: owner-mpi-21@XXXXXXXXXXXXX
Precedence: bulk
Reply-To: mpi-21@XXXXXXXXXXXXX

At 01:07 PM 5/10/2001 -0500, Andrew Lumsdaine wrote:
Hum. If I understand the idea of "collective" properly, then I don't see how
this can be accomplished without something like a barrier occurring.  Does
not the requirement that all the processes have to call the MPI routine also
imply that it is an error if they do not all do so?  Does not that also
imply that each process needs to wait until all processes have made the
collective call in order to know whether the collective call has been
properly made or not?

Otherwise, I don't think MPI_Finalize would be truly collective.  I.e., if
one can call it and receive indication of success regardless of whether
anyone else ever calls it or not, then it seems like a local procedure.

The behavior of erroneous programs is not defined.  You're right; a user-friendly implementation may want to *add* barrier semantics to help catch this error, but MPI makes no requirement that all (or even any) errors be detected (neither do most standards).

Note that MPI_Bcast( (void *)0, 0, MPI_INT, 0, MPI_COMM_WORLD ) is also a valid usage of collective routine but there is no requirement that prevents an implementation by turning this into a purely local no-op.

Bill