Hi Hannes,
This is excellent and very comprehensive work, thanks for sharing this.
With regards to the AMBER 12 vs AMBER 11 numbers, are these all constant
pressure runs (with isotropic or anisotropic pressure scaling?) - Some of
the code was reworked as part of adding the constant surface tension code
and this may account for some of the differences seen. I'd be interested to
know if the same differences are observed for NVT.
Also the Addendum mentions the MPI and compiler versions used for AMBER 12
but the original report (Mar 2012) doesn't give this info. Did you use the
exact same MPI and compiler versions for AMBER 11 as AMBER 12 in these
tests? That might also be causing a difference, especially if some of the
collectives over large core counts improved in a new MPI version.
One minor comment as well if I may. I know the standard approach used for
benchmarking in computer science is to plot log of execution time against
core count but I tend to find this very difficult to interpret since time is
an inverse scale and so the differences as you move to the right of the plot
become harder and harder to see. Can I suggest plotting throughput (ns/day)
against core count since this is then a linear scale and is much easier to
interpret. It is also what people really want to know, how much simulation
can I get done in a given time.
All the best
Ross
> -----Original Message-----
> From: Hannes Loeffler [mailto:Hannes.Loeffler.stfc.ac.uk]
> Sent: Thursday, July 05, 2012 5:55 AM
> To: amber.ambermd.org
> Subject: [AMBER] AMBER 12 benchmarks
>
> Hi,
>
> I finally found the time to do the AMBER 12 benchmarks on the iDataPlex
> and the Cray XE6. I wrote an addendum briefly summarising the results.
> The document can be found at
> http://www.stfc.ac.uk/CSE/randd/cbg/Benchmark/25241.aspx
>
> As usual the benchmark timings can be downloaded as well.
>
> Let me know if there are any questions or comments.
>
>
> Thank you,
> Hannes.
> --
> Scanned by iCritical.
>
> _______________________________________________
> AMBER mailing list
> AMBER.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Thu Jul 05 2012 - 09:30:03 PDT
