Newsgroups: comp.parallel.pvm
From: papadopo@cs.utk.edu (Philip Papadopoulos)
Subject: Re: Patch #8 for PVM version 3.3 is available on Netlib
Organization: CS Department, University of Tennessee, Knoxville
Date: 7 Aug 1995 09:31:17 -0400
Message-ID: <4054j5INNclh@duncan.cs.utk.edu>

In article <DCt3xC.n51@cs.dal.ca> edemaine@ug.cs.dal.ca (Erik Demaine) writes:
>weicheng jiang (wjiang@thud.cs.utk.edu) wrote:
>: The SP2 port in PVM3.3.8 is based on IBM's MPI-F, not MPICH. IBM's MPI-F
>: offers similar communication performance as MPL.
>: You must have IBM's MPI-F installed on your SP2 to take advantage of this 
>: special port.
>
>If the port is based on MPI-F and MPI-F is a _true_ MPI implementation, then
>why wouldn't it work atop MPICH?  MPICH has a special implementation for the
>SP-2 using MPL, and I have achieved _very_ good results.  Unfortunately, I do
>not have MPI-F to compare to, but I seem to recall a paper on this topic that
>realized commertial implementations are only slightly faster than the free,
>portable ones.
Weicheng aptly pointed out that MPI is *only a message passing interface*.
The MPI-1 guys punted on process control (you know, "spawning" and other things)
but MPI-2 is considering different process control proposals to include
in the next standard. The upshot is that starting MPI processes is different
for every version of MPI and this is where the really hard part comes
in for PVM.  Since MPI is static in nature, the pvmd can only be part of 
one MPI_COMM_WORLD. Consider 2 tasks spawned *separately* by a single pvmd
on an MPI-based machine:

           Time T=0             Time T=1
        _______________          -------
       | pvmd -- task1 |        | task2 |
        ---------------          -------
         MPI_COMM_WORLD         MPI_COMM_WORLD' 


The real trick is to get task1 and task2 to be able to send messages
back and fourth when the don't have a common frame of reference
(i.e their WORLDS are different).  Since Weicheng is the guy who did
the port, he can explain how he gets this all to work.

-Phil


