Newsgroups: comp.parallel.pvm
From: hzhou@cs.utk.edu (Honbo Zhou)
Subject: Re: Tasks communicating on the same workstation
Organization: CS Department, University of Tennessee, Knoxville
Date: 3 May 1995 14:22:54 -0400
Message-ID: <3o8hluINNmti@duncan.cs.utk.edu>

In article <3o6ur4$7qb@delphi.cs.ucla.edu> "Lawrence L. Chan" <lchan@cs.ucla.edu> writes:
>I have one more question on pvm's message delivery.  If I have 2 tasks running
>on the same workstation, and they communicate using pvm, will pvm be smart
>enough to use shared memory?  I'm worring about performance in here if pvm

PVM doesn't provide shared memory support for communication between local tasks
mostly because it doesn't worth the effort (I think). There is an effort to use
shared memory for the communication bewteen pvmd and local tasks (to alleviate
the traffic of pvmd) but it turned out no performance gain has been reached due
to reasons like slow semaphore and context switching (see the paper "Faster
Message Passing in PVM" at http://www.epm.ornl.gov/~zhou).

The communication between local tasks is supported by UNIX domain socket (instead
of Internet domain socket) in the current release of PVM. The UNIX domain socket
based communication is about 30% faster than the Internet domian socket based
communication.
 
>can't distinguish tasks that need network communication versus tasks that
>does not cross machine boundaries.  Will pvm treat all communicating tasks as
>if they were all remote?

Honbo


