[gmx-users] Re: gmx-users Digest, Vol 1, Issue 1430

Pim Schravendijk schraven at mpip-mainz.mpg.de
Tue May 18 21:04:05 CEST 2004


Thanks to the tip from Mike Sullivan (thanx!) to use the 386 rpm I was
able to do runs on a more plausible speed. I did one cpu runs on these 
systems:

Type	Ghz 	L2 Cache (I think this is L2)
---------------------------------------------
Mac	2.0 	512 kb
Xeon	2.8 	512 kb
Opteron 2.0 	1024 kb

And here are my results, I added the ones of David from November 2003

System	Mac(1)	Xeon(2)	Optn(3)	Optn(4)
-------------------------------------------
Villin	11220	9267	8807	9340
lzm	3048	2878	1801	2541
lzmpme	1438	1529	1281	1600
dppc	143 	151	173	185
polymer 3306	3945	3787	4192

(1) With espresso running on other node during most runs.
(2) Using the intel compiler.
(3) With two runs at one 2-cpu node for most runs; using the 386 rpm 
binary.
(4) Run with probably better compilation by David v.d. Spoel (gromacs 
developer).

I am a bit confused by these results, it seems that the Mac has some 
advantage from the altivec loop on the systems with a lot of water, but 
really loses when it comes to large amounts of non-water molecules. It 
could also be a memory related problem, but it that was the main cause the 
polymer should also do well on the Mac.

For the huge bilayer systems the opteron wins I would guess from these 
rough statistics, but the medium-sized systems might be handled better by 
a xeon or mac.

Do these data look plausible, or are there people with completely other 
experiences?

Greetings, Pim

--
Pim Schravendijk - PhD Student
Max Planck Institute for Polymer Research
http://www.mpip-mainz.mpg.de/~schraven/




More information about the gromacs.org_gmx-users mailing list