Newsgroups: comp.parallel.mpi,comp.parallel.pvm
From: wjiang@cs.utk.edu (weicheng jiang)
Subject: Re: MPI and PVM bandwidth comparisons on various networks
Organization: CS Department, University of Tennessee, Knoxville
Date: 15 Jul 1996 08:27:46 -0400
Message-ID: <4sddg2INN307@thud.cs.utk.edu>

In article <4rjfiv$1bc@sulawesi.lerc.nasa.gov> ciula@warhol.lerc.nasa.gov (Kim Ciula) writes:
>I recently compared the performance of several MPI implementations and
>PVM on various networks of our cluster of IBM RS/6000 workstations.  The
>results are on the web at:
>
>   http://www.lerc.nasa.gov/WWW/ACCL/PARALLEL/TIMING/benchmark.html
>
>I compared CHIMP 2.1.1c, LAM 6.0, MPICH 1.0.12, and PVM 3.3.11 performance
>on our ATM, FDDI, FCS, and ALLNODE networks.  I hope the results are of
>some interest.
>
>--
>Kim Ciula                         |  216-433-6596  Mail Stop 142-5
>Sterling Software at              |
>the NASA Lewis Research Center    | "Be excellent to one another"

Interesting. Here are the performance numbers I collected on a HP Fibre-
Channel Cluster, using the nntime.c example from the standard PVM 
distribution, and our enhanced version of PVM:

Latency					154 microseconds
Bandwidth (80K msg)		20 MegaBytes/second

The Fibre-Channel was quarter-speed, i.e., 25 MB/s specification.

Actual output from nntime:

Roundtrip T = 307 (us)  (0.0000 MB/s)      Data size: 0
Roundtrip T = 415 (us)  (0.3855 MB/s)      Data size: 80
Roundtrip T = 439 (us)  (3.6446 MB/s)      Data size: 800
Roundtrip T = 1297 (us)  (12.3362 MB/s)      Data size: 8000
Roundtrip T = 8065 (us)  (19.8388 MB/s)      Data size: 80000


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

