[gmx-users] Gromacs on Gigabit Ethernet Dual Opteron Cluster

[Bert de Groot]

Ghermes Chilov wrote:
> 
> Hi!
> 
> D> On Wed, 2003-10-01 at 18:53, Ghermes Chilov wrote:
> 
> >> I've got preliminary results of Gromacs's performance on Dual Opteron
> >> cluster (Gigabit Ethernet, LAM-MPI 7.0). The scalability is good (>80%)
> >> with up to 3 nodes used, but than it drops (44% for 8 nodes) and this
> >> decrease seems to correlate with CPU load (which also drops from
> >> 100-95% for 1 node to 60% for 8 nodes).
> >>
> >> Should I play with LAM compilation options or go fishing somewhere
> >> else? Probably parameters --with-tcp-short and --with-shm-short for
> >> Gigabit Ethernet may be increased? (with respect to 500 kb for Fast
> >> Ethernet).
> D> which system do you use?
> 
> D> I'd suggest starting with the DPPC benchmark. Don't forget to use
> D> -shuffle for grompp
> 
> AF> What system are you benchmarking with? It should be big enough,
> AF> like the 'dppc' from the benchmarks suite.
> 
> The system (dummified protein in water)is big enough:~ 107.000 atoms,
> -shuffle -sort were used.
> 

was this using PME or cut-off? I guess it was cut-off, otherwise I'd be
surprised about the scaling of 80% for three nodes...

If you used cut-off, would you mind repeating the benchmark with PME?

cheers,

Bert

____________________________________________________________________________
Dr. Bert de Groot

Max Planck Institute for Biophysical Chemistry
Computational biomolecular dynamics group
Am Fassberg 11 
37077 Goettingen, Germany

tel: +49-551-2011306, fax: +49-551-2011089

email: bgroot at gwdg.de
http://www.mpibpc.gwdg.de/abteilungen/071/bgroot
____________________________________________________________________________