AMBER Archive (2002)

Subject: Re: Scaling of Sander on Linux Clusters (fwd)

From: Stéphane Teletchéa (steletch_at_biomedicale.univ-paris5.fr)
Date: Tue Aug 27 2002 - 04:31:16 CDT

Le Mardi 27 Août 2002 10:56, Sanjeev B.S. a écrit :
> >The problem of network vs cpu is well known and documented but i think in
> >the case of sander in explicit water (PME on)
>
> But benchmarks on amber7 at amber site (though not available for high
> clock speed linux) shows good scaling, at least for 1.2GHZ AMD! It is
> confusing, indeed. Some benchmarking attempt has resulted equal speeds on
> 1.6GHz and 1GHz 4 processors giving very similar numbers. Looks like it is
> x86 based systems with higher clock speeds which are at disadvantage, I am
> not sure.

Available on the amber site ... I agree, but i have never been able to
reproduce their figures with my own systems !
Nevertheless, i'm thinking in terms of dozens of CPU (16, 32, 64, 128), so
even if it scales not too badly for 2-4 processors, i need speedup until
there. Without these kind of speedup, clusters are just an utopy.
Of course, you can benefit from the powerfull of one dual-node ...

Stef