Newsgroups: comp.parallel
From: Brian Wainscott <brian@lstc.com>
Subject: Re: [Q] which will be standard MPI or share memory
Organization: Livermore Software Technology Corporation
Date: Mon, 13 Feb 1995 17:42:53 GMT
Message-ID: <D3vGC6.8rK@lstc.com>

In message <3h69ei$bc5@news.cais.com> - duffy@cais2.cais.com (Duffy Men) writes
:
>A IBM engineer told me in the future the MPI (Message passing interface) 
>will be the standard in parallel programming.  Does anyone has option on 
>it?  Thank you.

We have a large application running on several machines: Paragon,
SP2, T3D, workstation networks, and other ports in the works.  I've
seen/used NX (Intel), EUIH (IBM), PVM, and others.

In my opinion (you asked for opinions!), MPI is significantly
better than all of the above.  In fact, it has a few features
that I had to supply myself on the other machines.
(reduction on a subgroup of the processors -- using MPI communicators it is 
trivial).  It is also the first widly used specification (the first I ever
heard of anyway) that will allow the development of good third party
libraries, the coupling of disparate codes, etc.  This is all possible
because the communicators are insulated from each other, so building
up pieces of large programs can be done without message passing conflicts.

In fact, our code (a large Finite Element Analysis code, about 300K lines
last time I checked) is now written in MPI, and runs on each of the 
above machines via translation libraries I wrote (MPI->NX, MPI->EUIH, 
etc.)Of course the libraries only supply those features I need.

We are using MPICH on the workstations.  I'm greatly looking forward to
true native MPI support on a wide range of hardware.  I expect it is
comming, because I also believe that MPI will be the standard.  At
least it SHOULD be.

--------------------------------------------------------------
Brian Wainscott
brian@lstc.com


