[prev in list] [next in list] [prev in thread] [next in thread] 

List:       beowulf
Subject:    Cobalt Benchmarks
From:       Luke Lonergan llonergan () hpti ! com
Date:       1999-10-28 2:30:31
[Download RAW message or body]

>     Cobalt does it's own partitioning using ParMETIS.  ParMETIS decomposes
> into perfectly (or extremely close) load balanced zones with
> minimum surface area contact.  Because ParMETIS also decomposes in

I wonder what ParMETIS is using. Sounds like what spectral bisection used to
do, but it cost a lot more in CPU time...

What I was referring to as "ill conditioned" partitioning is the kind of
thing that happens with models that have concavities or ductwork. Depending
on the partitioner, you can often end up with sliver zones that don't have
good surface area/volume characteristics. However, I think all that would do
is throw off your load balancing, and not reduce your message sizes.

Your results just seem too good to be true! Can you characterize the message
sizes? Greg Lindahl has a tool that you can link into your MPI based code
that will acquire statistics about the message sizes that your code uses
during its run. There are other such things out there, but this is really
simple and effective (he calls it "simple.c"). I would be very interested in
seeing what it says about COBALT on a real problem...

BTW, if COBALT runs this well on your Intel cluster, it would scream on our
Alpha EV6/Myrinet machine. I think 4x/CPU might be possible. Want some time
to benchmark on it?

Luke

-------------------------------------------------------------------
To unsubscribe send a message body containing "unsubscribe"
to beowulf-request@beowulf.org

[prev in list] [next in list] [prev in thread] [next in thread]