Cluster Benchmarks

[Andrew Shewmaker]

On Thu, 9 May 2002 08:24:13 -0400
"Eric Miller" <emiller at techskills.com> wrote:

> Another beginner question.  What are some of the widely-used benchmarks for
> MPI clusters?  I am running Scyld and would like to test out my cluster.(I
> think the Scyld distro comes with some basic benchmarks like linpack, but
> are there some others?)
> 
> Also, what are some programs that will really punish a cluster?  I would
> like to see my nodes get really hammered, I can currently get about 65%
> proc. usage, but I am just using the fractal renderer.

Have you looked through the mailing list archive?  An easy way to search it 
is to go to www.google.com and type "site:www.geocrawler.com beowulf benchmark" 
in the box.

Of course that returns absolutely tons of links.

Paralogic has a nice little package of benchmarks.  Most only involve one 
system.
http://www.plogic.com/bps

NetPIPE measures point to point, and it has one version written with MPI.  It 
is useful to see what overhead your implementation of MPI has.
http://www.scl.ameslab.gov/netpipe/

Every once in a while I see a HINT benchmark.  A company tried to commercialize 
it, but it doesn't seem to be devoloped any more.  The license in the tarball 
is GPL, I believe.  It uses a measurement called Quality Improvement Per Second 
(QUIPS) instead of something like MFLOPS.  I always liked it because it shows 
how performance drops across memory regions (cache, main, disk).
http://www.scl.ameslab.gov/HINT/

This paper by Dell used the HINT and NPB benchmarks.
http://www.dell.com/us/en/biz/topics/power_ps3q00-beowulf.htm

If you want to punish an individual cluster node, use cpuburn.
http://users.ev1.net/~redelm/

You could always just run another instance of the fractal render.

-Andrew