HOWTO discriminate switches?

Richard Walsh rbw at
Mon Sep 17 07:02:51 PDT 2001

Chris Wilson wrote:

>The situation is that we are trying to decide on the right switch for 
>use in a 80 node cluster. The first option is to use a single switch with 
>160+ fast ethernet ports. The second is to use eight 10 port gigabit 
>switches plus the current gigabit switch for routing.
>The codes to be run on the cluster are a mixture of the embarassing
>parallel, bandwidth hog and latency sensitive -- which complicates
>matters. :)
>What I need is guidance on how to measure (and indeed what to be
>measuring!) the switches to discriminate between the options. Presumably
>netperf is a good starting point, but pointers to resources on switch
>comparisions would be useful.

Another alternative which starts to make sense as the switching cost
of your system rises as a percentage of total cost is a switchless
system in the form of a 1D or 2D torus using SCI cards which have 
extremely low latency and high bandwdth. Take a look at the cluster
at the University of Delaware based on this design. If you do the 
math, I think you will find that for larger clusters which would 
require multiple layers of switches or a "mega-switch" to build
in reasonable bandwidth, the SCI torus design is no more costly and
has better performance features (latency and bandwidth). SCI ratings
are equal to or better than some custom engineered networks from 
vendors of SMP and MPP systems (Cray, Sun, SGI).


# Richard Walsh
# Project Manager, Cluster Computing, Computational
#                  Chemistry and Finance
# netASPx, Inc.
# 1200 Washington Ave. So.
# Minneapolis, MN 55415
# VOX:    612-337-3467
# FAX:    612-337-3400
# EMAIL:  rbw at, richard.walsh at
# "What you can do, or dream you can, begin it;
#  Boldness has genius, power, and magic in it."
#                                  -Goethe
# "Without mystery, there can be no authority."
#                                  -Charles DeGaulle
# "Why waste time learning when ignornace is 
#  instantaneous?"                 -Thomas Hobbes

More information about the Beowulf mailing list