There should be a bit of a speedup from 8 to 9, though the biggest impacts
in 9 are in high scaling, and you are not doing high scaling at 16 x 2 cpu.
I would guess somewhere around 10-30% for going from 8 to 9, depending on
various factors, in this range; part of the problem in making good
comparisons is a lot of the amber 8 benchmarks are on machines we no longer
use. But 9 was overall a substantial performance boost over 8, but once
again, it's the scaling that is more noticeable. For amber 10, I have not
gotten around to official benchmarks yet, and while we got some speedup, it
is not as good as going from 8 to 9. So for single processors in 10, you
can get up to 30% speedup with the nve ensemble, less than this for nvt, and
practically no gain for the npt ensemble - these are single and low
processor count numbers. Then there are fft bottleneck issues at high
scaling that were fixed, that in some scenarios can give you roughly a 50%
speedup - but this is only at the high end. I have actually not run pmemd
10 on opterons myself as of yet, so the speedup estimates here are generic.
The best way to benchmark systems is to use the benchmarks we ship - then
there is some standard for comparison. I personally still like jac and the
factor ix benchmarks myself; some folks do feel that at 23k atoms the
relevance of jac is subsiding. The factor ix benchmark as shipped with
amber does not print trajectory; it is more representative to perhaps do the
nvt and npt benchmarks I publish with factor ix. Also, most folks are more
interested in 2 fsec timesteps these days, apparently.
Regards - Bob
----- Original Message -----
From: "Ed Pate" <pate.math.wsu.edu>
To: <amber.scripps.edu>
Sent: Friday, June 20, 2008 1:08 PM
Subject: AMBER: Amber 10 timing
> Dear Amber community:
>
> We have recently installed Amber 10 as an upgrade from Amber 8. As a test
> case for timings, we ran a pmemd simulation of a protein in a box of TIP3P
> waters. The entire simulation was approximately 100,000 atoms. I have
> been unable to find any timing benchmarks for Amber 10. We find about a
> 10% speedup with Amber 10 over Amber 8. Is this what we should expect?
>
> The system nodes are dual processor, dual core Opteron 2220s, infiniband
> interconnects, Suse 10.0. Amber 10 and Amber 8 were compiled with
> Portland Group compilers. The pmemd job was run on 16 nodes.
>
> Thanks for the help.
>
> Ed Pate
>
> -----------------------------------------------------------------------
> The AMBER Mail Reflector
> To post, send mail to amber.scripps.edu
> To unsubscribe, send "unsubscribe amber" (in the *body* of the email)
> to majordomo.scripps.edu
>
-----------------------------------------------------------------------
The AMBER Mail Reflector
To post, send mail to amber.scripps.edu
To unsubscribe, send "unsubscribe amber" (in the *body* of the email)
to majordomo.scripps.edu
Received on Sun Jun 22 2008 - 06:08:02 PDT
