Newsgroups: comp.parallel.pvm
From: Nick Nystrom <nystrom@psc.edu>
Subject: Re: PVM speed efficiency
Organization: Pittsburgh Supercomputing Center
Date: Wed, 26 Feb 1997 09:11:07 -0500
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Transfer-Encoding: 7bit
Message-ID: <3314447B.41C6@psc.edu>

Nick Maclaren wrote:
> [...]
> And, to cap it all, PVM is both a 'standard' and an implementation,
> and the implementation is built on top of TCP/IP.  Even on machines
> with a hardware latency of microseconds, this can lead to latencies
> of milliseconds.

This statement is somewhat misleading since vendors routinely offer
implementations of PVM tuned for specific architectures.  For
example, on the Cray T3D, with a shared-memory "put" latency of
1-2 microseconds, pvm_psend/pvm_precv latencies are only about
25-30 microseconds.  Similarly, DEC's implementation of PVM for
their Memory Channel systems realizes lower latencies than would
be possible using TCP/IP.  TCP/IP is often the common denominator
for connecting heterogeneous systems, but I think it's fair to say
that many PVM implementations for high performance systems use
more streamlined protocols tuned to those architectures.

Nick