Newsgroups: comp.parallel.mpi
From: bijal@nuc.berkeley.edu (Bijal C. Modi)
Subject: Re: MPICH : MPI rank order  mapping w.r.t process group file.
Organization: University of California
Date: 11 Apr 1996 04:45:40 GMT
Message-ID: <4ki2pk$fhu@agate.berkeley.edu>

In article <4kh54r$1frh@watnews1.watson.ibm.com> you write:
>I have come across a one is to one mapping of the rank with respect to the 
>order of the node names /host names mentioned in the proc group file.
>For example  if the proc group file *.pg has the follwoing entries
>
>host1.domain1   0  $HOME/MPI/a.out
>host2.domain2   1  $HOME/MPI/a.out
>host3.domain3   2  $HOME/MPI/a.out

This means h1.d1 would have one copy of a.out (and that no other process can be
spawned) , h2.d2 would have one copy of a.out and h3.d3 would have 2. So total
of 4 processes distributed over 3 hosts.

>
>the ranks of host1 is 0, host2 is 1 and host3 is 3. If I interchange host2 and
                                                 ^^^  
>host 3 the ranks also follow suit. 

It should be 2, 3?

>I am not sure whether this is a part of MPI standard or is it the underlying
>p4 layer implementation specific??
> Would an MPI implementation for SP2 exhibit similar behaviour?
>I ma trying to port MPI on to a communication layer of a program.The program
>sends data by using a "node id" which is translated by the comm layer to the
>corresponding "socket". How do I get to know which node has what rank at
>runtime?
>-regards

MPI standard doesn't say anything about procgroup file, so obviously it's
not a part of MPI standard. It's a p4 layer implementation specific
feature. You can always use MPI_Get_processor_name to find out which
machine/node has which processes with given rank.

-Bijal
 bijal@fusion.berkeley.edu
--
                *      /\                       *
        *             |()|        *         Bijal Modi                    *
           *          |  |                  bijal@fusion.berkeley.edu
 *              *     |  |     *        *   U.C. Berkeley

