gige benchmark performance

[Velocet]

On Fri, Mar 15, 2002 at 04:51:02PM -0800, Tim Carlson's all...
> On Thu, 14 Mar 2002, Mark Hartner wrote:
> 
> > We are working on putting together a 32 node cluster at the University of
> > Utah. I have benchmarked several Gigabit cards over cat5,
> > Netgear GA620T
> > Netgear GA622T
> > Intel Pro/1000T Server
> 
> I put together some similar numbers at have posted them here:
> 
> http://www.santafe.edu/~tim/netpipe.html

So what do all these numbers really mean? As we all see often, "thats just
a benchmark, doesnt relate to the real world of course". It does in some
way, but it needs to be specified.

I am not sure how many people use which of the various parallel message
interfaces between nodes, but it would be good if we could somehow get
figures that relate to these packages.

I personally am interested in LAM/MPI's performance with Gromacs. For the
limited set of jobs I've played with, I've only seem LAM want to use
packets of 700-1000 bytes. At that size we're not interested in throughput
of course, at least not directly - we're more interested in latency. I am
not sure if the netperf numbers for each of the different configurations
we've seen in the last few days can predict what the LAM/MPI performance
will be for my types of applications or for the myriad of other uses
that people may want to put GBE networks to.

Is there a way to relate these numbers to LAM/MPI or other parallel meshes
that beowulfers use? Is there a standard network benchmark that should be
used for LAM?

Or are we reduced to running exactly our particular jobs of interest on
the exact configuration we'd use in production in order to get relevant
stats? Thats not so useful when trying to spec out a new cluster for
purchase - being able to relate real world performance to simpler benchmarks
would obviously be very useful to all of us. But how to go about doing
this? 

(Do we want to make a standard beowulf-l benchmark package that runs a subset
of benchmarks that may shed some light on all the different performance
concerns that various factions on this list have? :)

/kc


> 
> Tim
> 
> Tim Carlson
> Voice: (509) 376 3423
> Email: Tim.Carlson at pnl.gov
> EMSL UNIX System Support
> 
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org
> To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf

-- 
Ken Chase, math at velocet.ca  *  Velocet Communications Inc.  *  Toronto, CANADA