Path: utzoo!utgpu!news-server.csri.toronto.edu!cs.utexas.edu!uunet!samsung!emory!hubcap!linc
From: linc@cs.washington.edu (Calvin Lin)
Newsgroups: comp.parallel
Subject: Re: Is message passing more efficient than shared mem ?
Message-ID: <12671@hubcap.clemson.edu>
Date: 17 Jan 91 22:22:42 GMT
Sender: fpst@hubcap.clemson.edu
Organization: U of Washington, Computer Science, Seattle
Lines: 41
Approved: parallel@hubcap.clemson.edu

>From: roland@sics.se (Roland Karlsson)

> Hi Gerhard,
>
> > In the Proceedings of the ICPP 1990 I found a very interesting paper
> > suggesting that programs based on message passing are more efficient
> > than those based on shared memory - even on shared memory
> > architectures.

> > Is anyone working on this ? Are there any intermediate results yet ?

     As far as we know, there has not been any further work in this
area.

>
> I hope that neither you nor the authors in the article get offended,
> but this is nonsense.  I assume you have simplified the results from
> the article.  You can not compare two different ways of implementation
> without knowing what to implement.  I have seen several examples of
> programs that have been improved when moved from a loosely coupled
> machine to a shared memory machine.  The usual trick is to replace
> message passing with direct access of other processes memory.  There

     No, we're not offended, but we'd like to point out that what you've
stated does not contradict our work.  Our paper does not compare the
performance of a message-passing program on a loosely coupled machine
against the performance of a similar program on a shared memory machine.
Rather, we are interested in comparing the performance of a program 
written in the non-shared memory model with that of a shared memory
program, both executing on the same machine.

> ... One more
> example: I am working on a very efficient implementation of
> Or-parallel Prolog.  There is no way of doing this, with the same
> efficiency, with message passing.  ...

     Just to set the record straight, our paper does not claim that 
any one model is superior for all applications.

			Calvin Lin
			linc@cs.washington.edu