From: treumann@XXXXXXXXXX
X-Lotus-FromDomain: IBMUS
To: William Gropp <gropp@XXXXXXXXXXX>
cc: mpi-core@XXXXXXXXXXX
Date: Thu, 27 Jan 2000 12:58:20 -0500
Subject: Re: MPI_PROC_NULL in MPI_GROUP_TRANSLATE_RANKS
Mime-Version: 1.0
Content-type: text/plain; charset=us-ascii
Content-Disposition: inline

Bill suggested clarifying MPI to allow MPI_GROUP_TRANSLATE_RANKS to accept
MPI_PROC_NULL and translate it to MPI_PROC_NULL.

I have taken a closer look at the MPI-1 & 2 and found little explicit
characterization of MPI_PROC_NULL.  In Section 3.11 it is characterized as
a "dummy" source or destination.    A paraphrased definition of  source or
destination is  " a rank in a communicator group".   This can be read as
implying, MPI_PROC_NULL is "a dummy rank in any communicator group"

No where in MPI-1 or MPI-2 that I can find do we characterize it as an
"invalid" rank.  Also, no where to we identify a usage for  it as
prohibited.   I will now make the stronger assertion that we ought to add
this clarification of the standard because it is logical, useful and there
is nothing to imply it ought not to  be allowed.

I do not see the need to allow MPI_UNDEFINED. The standard does not even
hint that MPI_UNDEFINED is like a rank.  After all, MPI_GET_COUNT also can
return MPI_UNDEFINED.

            Dick

BTW -- How do clarifications like this become "official" and do we have a
place they are documented?



Dick Treumann                IBM RS/6000 SP Development
Tele (914) 433-7846         Fax (914) 433-8363