Updated (and better) compiler flags are also a possible explanation (I
think I recall Robin adding -mtune=native to the GNU compiler optimization
flags between the Amber 11 and Amber 12 releases). Unless care was taken
to apply exactly the same compiler flags (and obviously compilers) from
Amber 11 to Amber 12, I'm not exactly sure where to ascribe the performance
differences to.
Are the reported benchmarks "out-of-the-box" compilations (sans edits to
the config.h file)?
On Thu, Jul 5, 2012 at 11:49 AM, Hannes Loeffler <Hannes.Loeffler.stfc.ac.uk
> wrote:
> On Thu, 5 Jul 2012 15:06:09 +0200
> Jan-Philip Gehrcke <jgehrcke.googlemail.com> wrote:
>
> > Hey,
> >
> > interesting work.
> >
> > Are there any changes in the "core MD engine" (memory management,
> > message passing, ...) of pmemd from Amber 11 to Amber 12 that support
> > the measured differences with respect to performance and memory
> > consumption or are these differences due to slight variances in the
> > experimental environment (due to other jobs running on the same
> > nodes, varying network performance, ...)?
>
>
> Slight variance would be a difference of a few percent and we see that
> in test runs. But the differences on the Cray XE6 with the larger
> systems are indeed significant. I do not know why that is but I could
> reproduce those timings with a few additional runs.
> --
> Scanned by iCritical.
>
> _______________________________________________
> AMBER mailing list
> AMBER.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber
>
--
Jason M. Swails
Quantum Theory Project,
University of Florida
Ph.D. Candidate
352-392-4032
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Thu Jul 05 2012 - 10:30:02 PDT
