[Wrf-users] WRF Benchmark Results

[Brent L Shaw]

Hello all,

 

I am running WRF on a new AMD Opteron-based Linux cluster (using PGI
compilers, DM Parallel, RSL_LITE).  I have used similarly configured
clusters with great success, but on this particular system I am seeing
very poor scalability when exceeding more than two nodes (I actually see
0 and even some negative scalability for a problem that should be
scaling well across the cluster).  The two main differences between this
new cluster and previous clusters I have used are:

 

1.	The compute nodes on this new cluster have two dual-core Opteron
processors, whereas in the past I have only had access to dual
single-core Opterons.
2.	This compute node only has Gigabit Ethernet for MPICH, whereas
in the past I have always had Myrinet.  The "conventional wisdom" is
that for a small cluster (8 nodes or less), there is not a huge
advantage in using Myrinet vs. Gig-E.  However, I wonder if that is
still true now that we can run 4 processes/node instead of 2?  

 

Do any of you out there in WRF-land have a similar cluster for which you
would be willing to share some benchmark results?   I am in the process
of running the standard WRF CONUS 12-km benchmark on a range of 1-16
processors.

 

Thanks in advance to anyone willing to discuss this with me!

 

Brent

 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mailman.ucar.edu/pipermail/wrf-users/attachments/20070807/521c4371/attachment.html