[mpich2-dev] Quad core Harpertown and mpich2

[Hee Il Kim]

Hi,

I recently made a Rocks cluster consisting of 10 computing nodes with dual
Harpertown (E5420) and 8 giga memory per node. The computing nodes are
connected by a gigabit ethernet network.

I tested benchmarks with mpich2 builded with Intel compilers (non-commercial
version 10.1) and found the scalability is not good. Actually it was bad. I
expected much better results because even though I used the gigabit network,
the benchmark problem doesn't require big communication between the cpu
cores. I used Cactus benchmark, program for numerical relativity. Similar
results reported from my collegue who are using Gadget-2, a kind of N-body
simulation code.

At first I configured mpich2 with ssm channel. But after reading some
documents on nemesis and multicore CPUs like Clovertown, I tested mpich2
with nemesis. Unfortunately I got worse results than with ssm.

I'm a newbie on this field. Due to the lack of fund, I can't upgrade the
network to low-latency networks for some time. Is there any other way to
improve the scalability of my cluster? What's the best configuration of
mpich2 for this Harpertown cluster? I got similar results with openmpi and
mpich too.

I appreciate your helps in advance.

Regards,

Hee Il Kim

-- 
Dr. Kim, Hee Il
Astronomy Program
Dept. of Physics and Astronomy
Seoul National University, Seoul, 151-747, Korea
emails: khi at astro.snu.ac.kr, heeilkim at gmail.com
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.mcs.anl.gov/mailman/private/mpich2-dev/attachments/20080327/cf829def/attachment.htm>