[gmx-users] Clovertown vs Woodcrest gromacs experiences?
lindahl at cbr.su.se
Fri Aug 24 09:56:32 CEST 2007
Quad-core is the way to go. I recently benchmarked the scaling of the
CVS version, and with 8 independent jobs we get 85-97% throughput
scaling, depending on the type of simulation. And you get essentially
the same performance if you run two jobs using 4 cores each.
Even for a single simulation parallelizing over all 8 cores (lots of
communication), the scaling is 70-90% with CVS.
However, unless you're in a rush you might want to wait until
November when Intel is rumored to ship Harpertown (Penryn quad-core)
with more cache, higher frequencies, and lower price. AMD will also
release Barcelona (true quad-cores) in September, with higher
frequencies expected in october/november.
On Aug 24, 2007, at 12:53 AM, Alan Chen wrote:
> If anyone has some gromacs benchmarks numbers they'd be willing to
> share from clovertown
> quad cores. .. we're having to decide between buying 3.0 ghz DPDC
> woodcrest nodes (which
> we already have several of and are quite happy with) vs 2.66 ghz DP/
> QC clovertowns for the same price.
> I haven't seen any clovertown results in the benchmark page or in
> past listserv messages. .
> Obviously, we don't expect the clovertowns to scale perfectly for
> single jobs due to the architecture, but I'm very
> curious how they do with multiple jobs on the same node - we tend
> to run lots of independent trajectories of medium
> (~100,000 atom) sized systems in our lab.
> If anyone can share their experiences with gromacs on clovertown
> XEON systems it would be greatly appreciated!
> Alan Chen
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the gromacs.org_gmx-users