[gmx-users] GROMACS scaling at 64 cpus

[Berk Hess]

Hi,

This of course depends very much on all your settings.
Up to now I have found that a 3:1 ratio is optimal for rectangular boxes,
whereas 2:1 is optimal for rhombic dodecahedra.
64 cores was the limit for the benchmark system on that hardware,
so I had to increase the number of PME beyond the 1/3 to get optimal performance.

Berk


Hi,

I recently got a nice Sun X6275 blade system 
http://nf.nci.org.au/facilities/vayu/hardware.php (ok, not just for me!) 
whose hardware ought to be better, but very comparable to that used in 
the GROMACS 4 paper (http://pubs.acs.org/doi/abs/10.1021/ct700301q). 
Figure 5 there demonstrates good scaling out to 38 cpus, but it drops 
away from linearity by 64. On the same test system on this hardware (pdf 
attached; blue arrow is ideal linearity) the drop-off is still about the 
same.

For 64 cpus, that paper reports using 28 PME nodes, which is an 
appreciably lower ratio than the 2:1 for the rest of the tests. On my 
hardware I found that that choice cost about 30% performance, compared 
with a more reasonable use of 19 PME nodes (64 vs 86 ns/day, pdf 
attached). Is that 28 a typo for 18?

Mark
 		 	   		   		 	   		  
_________________________________________________________________
What can you do with the new Windows Live? Find out
http://www.microsoft.com/windows/windowslive/default.aspx
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://maillist.sys.kth.se/pipermail/gromacs.org_gmx-users/attachments/20091016/409efd0d/attachment.html>