Newsgroups: comp.parallel.pvm
From: wjiang@apollo.hp.com (Weicheng Jiang)
Subject: Re: HPPAMP with shared memory and TCP/IP
Organization: Hewlett-Packard Company, Chelmsford, MA
Date: Sat, 27 Jul 1996 16:42:04 GMT
Message-ID: <Dv7nq5.n2F@apollo.hp.com>

In article <31F8018B.6201@laplace.csb.yale.edu> Paul Adams <paul@laplace.csb.yal
e.edu> writes:
>  I observe very different performance using shared memory and TCP/IP
>with the HPPAMP architecture. A program run on 2 CPUs within the same
>machine (ie. using shared memory) takes ~250 seconds per CPU, while the
>same program run on 2 CPUs in different machines (both with HPPAMP
>architecture) via an ATM switch (ie. using TCP/IP) takes ~320 seconds
>per CPU.

This is to be expected, because shared-memory communication is much
faster than ATM. On our Fiber-channel cluster, the performance of PVM
is as follows:

			Shared-memory	Fiber-channel
Latency (us)	      	      7		    152
Bandwidth (MB/s)     	     154	     20

The performance over ATM with TCP is probably much worse than Fiber-channel.
(ATM data rate: 155Mbps, Fiber-channel: 266Mbps.)

--
Weicheng Jiang, Ph.D.            Chelmsford System Software Lab.
PVM Specialist                   Network Computing Division
Tel: 508-436-4613                Hewlett-Packard Company

