[gmx-users] .log file documentaion
nmoore at physics.umn.edu
Fri Jun 10 22:45:27 CEST 2005
>> I'm guessing that the 3rd column (homenr) specifies the number of atoms
>> initially assigned to each processor - is that correct? A reference
>> giving general summary of the .log output would be wonderful.
> Your guess is correct. Don't expect any documentation for this stuff. If
> something we could add some output in specific places (read: extremely
> obscure output). But since you geussed what it means, it doesn't seem to
> be a big problem...
> If you put this system on 512 nodes you will have an integer number
> (238) atoms per processor! Have you tried scaling the dppc benchmark?
Yeah, I've got some initial results for scaling that I'm looking at right
now, although only up to np=256 (havn't recompiled with MAXNODES=1024
yet). I can share them if you're interested. At present the scaling
decays rapidly beyond np =~48 or so, with maximum ps/d at np=128 of ~50k.
On one node I get about 21 ps/day; if np=2, 81 ps/day; if np=4,
320ps/day; These initial results are from only ns=50, rather than the
GROAMCS benchmark's suggested 5000 steps.
I've tried a bunch of compiler optimizations, as well as the -sort and
-shuffle grompp flags (which certainly improve performance). If you have
any suggestions for improving performance, I'd love to know.
More information about the gmx-users