[gmx-users] scaling summary

[Lynne E. Bilston]

Hi all,

I thought that I'd summarise what I found while trying to improve the 
scaling of jobs on my small cluster of dual athlons, in case anyone else 
has a similar setup.

Thanks to David for his suggestions on speeding up my scaling by 
recompiling lam with the short message option. That turned out to be the 
key. For the curious, I recompiled lam using the following options:

./configure --prefix=/directory --with-tcp-short=524288 --with-rpi=usysv

It increased my scaling on the dppc benchmark from 76% to 100%, according 
to the logfile. More meaningfully, it means that the dppc benchmark took 
1hr to run over two dual cpu machines, instead of 1hr 43 minutes without 
the short message option. Definitely worth the time to recompile lam!

Cheers,
-Lynne
________________________________________________________________________________
Lynne E. Bilston, PhD
Senior Scientist, Prince of Wales Medical Research Institute, and
Conjoint Associate Professor, Faculty of Medicine, University of New South 
Wales
Barker St, Randwick, NSW 2031 Australia
Tel: +61-2-9382-7924	Fax: +61-2-9382-2643
________________________________________________________________________________