[gmx-users] non-Myrinet Gromacs

[Anton Feenstra]

David van der Spoel wrote:

> On Mon, 2005-05-02 at 20:52 +0800, Yang Ye wrote:
> 
>>Shall be "doesn't scale well"
> 
> there are actually some (old) scaling benchmarks on the website with and
> without a fast network.

As CPU's grow faster, networks (esp. latency) has lagged behind. This is 
the big bottleneck for Gromacs scaling performance. On the same tack, 
improving the Gromacs serial performance (read, Eriks assembly loops) 
also worsens the scaling.

To give you a comparison, I've been running some Amber and Gromos runs 
lately. Even their 'optimized' versions run about 10 (ten!) times slower 
than Gromacs. So what 'we' can do on one (1) dual CPU machine, 'they' 
need at least 10 dual machines for (given perfect scaling, that is).
If you work out the time spent in network latency versus that of one 
integration step, and take a factor of 10 difference in the latter, you 
can see where the scaling goes... ;-)

-- 
Groetjes,

Anton

* NOTE: New Phone & Fax numbers (below) *

  _____________ _______________________________________________________
|             |                                                       |
|  _   _  ___,| K. Anton Feenstra                                     |
| / \ / \'| | | Dept. of Pharmacochem. - Vrije Universiteit Amsterdam |
|(   |   )| | | De Boelelaan 1083 - 1081 HV Amsterdam - Netherlands   |
| \_/ \_/ | | | Tel: +31 20 59 87608 - Fax: +31 20 59 87610           |
|             | Feenstra at few.vu.nl - www.few.vu.nl/~feenstra/         |
|             | "If You See Me Getting High, Knock Me Down" (RHCP)    |
|_____________|_______________________________________________________|