Newsgroups: comp.parallel.pvm
From: papadopo@cs.utk.edu (Philip Papadopoulos)
Subject: Re: COW PVM Question
Organization: CS Department, University of Tennessee, Knoxville
Date: 7 Aug 1995 09:16:14 -0400
Message-ID: <4053muINNbt3@duncan.cs.utk.edu>

In article <400vc5$omk@newshound.uidaho.edu> gree9031@buzzard.csrv.uidaho.edu (Greenup Cade) writes:
>
>I've been working with a program that was converted from a Mas-Par (with 
>a ton of processors) to PVM on a HP COW.  The clusted only has about 8 
>workstations in it.  The PVM program has over a hundred processes in it, 
>thus each workstation will have about 10 of them running on it.  The 
>question is when the processes within a workstation are exchanging data, 
>do they still put the information out on the network, just as if the 
>processes were on different nodes, or is PVM smart enough to keep the 
>communication internal to that single processor?
It uses TCP or UNIX domain sockets -- Defaulted routed packets will
go task-daemon-task.  PVM supports native communication paths for
shared memory machines (eg. SGIMP, Convex Exemplar, SUNMP, RS6KMP, ALPHAMP,
CRAYSMP) and message passing (e.g. PGON, I860, CM5).  The packets
don't go on the physical wire, since most (all?) machines understand
packets addressed to themselves. Still, the TCP/UNIX protocol is
pretty heavy-weight and sucks down the cycles.  

-Phil



