Newsgroups: comp.parallel.mpi,comp.parallel.pvm
From: Steve Huss-Lederman <lederman@breeze.cs.wisc.edu>
Subject: Re: PVM and MPI
Organization: CS Department, University of Wisconsin
Date: 02 Feb 1998 10:56:54 -0600
Message-ID: <dwpn2ga3pyh.fsf@breeze.cs.wisc.edu>

> "Nothing in the MPI standard describes cooperation across heterogeneous
> networks and architectures. [...] None of the existing MPI
> implementations can interoperate."

I'll risk trying to add some information on this topic without trying
to be biased to PVM or MPI.  The MPI standard was carefully designed
not to specify what protocols, etc. are used in transmitting messages.
This was deliberately done to allow each implementation to use
whatever was optimal on its platform(s).  As a result, different MPI
implementations use different mechanisms (transmission protocols,
headers, etc.) and they generally cannot send messages to each other
at this time.  Thus, you don't expect MPICH on machine A to be able to
send messages to LAM on machine B.  The above statement is (I believe)
refering to this fact.  However, this does not mean you cannot get
machine A to send messages to machine B.  You need to find an
implementation that supports both of the machines types.  MPI is very
clear that if and implementation supports two architectures then the
user can transparently send messages between these two types of
machines.  Thus, MPICH or LAM running on both machines can send
messages between machine A and machine B.  One negative that this can
cause is if the single implementation you need to use is not optimal
on one of the machines then message passing internal to that machine
can be slower.  I'm not saying this is necessarily the case but it
does hinder you from chosing different implementations on different
machines.  Finally, I'll note that MPI does not exclude different
implementations being able to communicate.  There is at least one
effort which is trying to help standardize how this would be done.  An
important point here is to do it carefully so as not to slow down
internal communications and allow for high speed between different
implementatons.  The MPI Forum deliberately chose not to set up these
mechanisms.  The main reason (as I recall from the discussion) was
that this was best done by the implemenators and not the Forum.

Hope this helps.

Steve