[gmx-users] GROMACS not scaling well with Core4 Quad technology CPUs

[Trevor Marshall]

Thanks to Yang Ye, Erik, Mark, I appreciate your help a few days ago.

I found a clear answer to the communication congestion in the citation 
given by Yang Ye, to Carsten Kutzner's Powerpoint presentation at URL
http://wwwuser.gwdg.de/%7Eckutzne/talks/kutzner04talk-gromacs.pdf

Page 21 has a slide showing the speedup versus the number of CPUs for 
several computer systems, clearly showing the congestion when the 
interconnect bandwidth is not very high. This graph needs to go somewhere 
on the Gromacs benchmarking pages, IMO.

I subsequently changed the Quad Core2 machine to use MPICH2 with gcc, so I 
could try out their new ch3:nemesis interconnect. The Q6600 CPU is running 
(overclocked) at 3GHz with a 1333MHZ FSB.  Using the same molecular model 
as before, I now consistently get 6.98 hours/nsec, not far from the 
theoretical limit of 6.1 calculated by Mark.  Interestingly, the 
performance is identical, whether I use np=3 or np=4, but the average load 
per CPU drops to 75% with np=4.  The load sharing under MPICH2 seems 
excellent.  I need to compute with a few different models to get a better 
handle on where any limitations remain.

Page 10 of Carsten's presentation suggested changing PME to 6 and 
fourierspacing to 0.178, which improved things by a few percent.

I am loath to adopt the CVS build because
1. it has been several years since I worked with repositories, and my 
memory is sketchy, and
2. I am using the Rotational-Constraints-patched version of mdrun 3.3.1 to 
allow longer runs in an NDLP environment, and am unsure whether the patches 
would be applicable to the CVS code.

Anyway, I have a system now which I can use to give pretty good throughput 
with the models I use most frequently, so I will refocus my efforts on the 
molecular modelling and away from the computer architectures :) Thanks 
again to all who helped, and please contact me at any time if you think of 
any other tips and tricks :)

Sincerely
Trevor

------------------------------------------
Trevor G Marshall, PhD
Director, Autoimmunity Research Foundation, Thousand Oaks, California
School of Biological Sciences and Biotechnology, Murdoch University, 
Western Australia
Patron, Australian Autoimmunity Foundation.
------------------------------------------