Newsgroups: comp.parallel.pvm
From: Paul Adams <paul@laplace.csb.yale.edu>
Subject: Re: HPPAMP with shared memory and TCP/IP
Organization: Yale University
Date: Mon, 29 Jul 1996 17:59:47 -0400
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Transfer-Encoding: 7bit
Message-ID: <31FD3453.41C6@laplace.csb.yale.edu>

Hello,

  thanks to Weicheng Jiang for the numbers re shared memory and Fibre
channel performance. However, I should have been more specific about my
observations. I have benchmarked on two sets of machines: 2 single CPU
HPs connected by ATM, and 2 multi-CPU HPs connected by ATM. If I use
shared memory on the multi-CPU machines or TCP/IP through the ATM on the
single CPU machines I get similar performance and parallel speedup (~1.9
for 2 CPUs). However, when using the ATM between the 2 multi-CPU
machines I get much worse performance and speedup (~1.5 for 2 CPUs).
Obviously, I expect shared memory to be the fastest thing possible (and
I do see slightly better performance), but why the large difference in
performance between the ATM(TCP/IP) on the single CPU machines and ATM
on the multi-CPUs? Does the shared memory version of PVM do something
really inefficient on the multi-CPU HPs when going over the network?
Going through TCP/IP on the multi-CPU machines seems to cost me ~70
seconds of CPU (20% of total CPU time). Any ideas would be appreciated.

  Cheers,
	Paul

-- 
========================================================================
| Paul Adams               Dept. Molecular Biophysics and Biochemistry |
| Postdoctoral Associate   Bass Center, 266 Whitney Avenue             |
| Office: 203-432-5066     Yale University                             |
| FAX:    203-432-6946     New Haven, CT 06511, USA.                   |
| http://xplor.csb.yale.edu/~paul   mailto://paul@laplace.csb.yale.edu |
========================================================================

