[gmx-users] Comparing Gromacs versions

[Mark Abraham]

On Fri, May 17, 2013 at 11:07 AM, Djurre de Jong-Bruinink <
djurredejong at yahoo.com> wrote:

> Dear all,
>
> We recently did some comparisons of gmx 4.5.5, 4.6.1+group and
> 4.6.1+verlet cutoff scheme. The system was a protein (amber99) in water
> (tip3p) in a triclinic box (23k atoms total). We're using pme with a
> realspace cutoff=0.9nm, gridspacing=0.12nm, npme=-1, which based on
> g_tune_pme gives (close to) best performance. Below some performance
> measure (ns/day) from 100,000 step runs:
>
> CPUs     G455        G461+group               G461+verlet
> 24            34.9          43.1                       35.3
> 36            47.6          57.8                       49.7
> 48            50.8          65.9                       60.4
> 60            59.9          81.8                       71.1
> 72            57.7          78.7                       78.2
>
> The scaling to more CPUs in 4.6.1 is much better (thank you developers for
> putting up all the hard work!), however I would have expected more
> improvements from the verlet scheme.


You are doing simulations with a lot of water (and perhaps with charge
groups), and that is the case where an unbuffered group scheme has the best
performance. How much you like the physics is another story. See
http://www.gromacs.org/Documentation/Acceleration_and_parallelization and
http://www.gromacs.org/Documentation/Cut-off_schemes

Note that both 455 and 461 have been compiled with MPI only (no OMP, GPU),
> and 461 was compiled with gcc (4.7.2). I did change the nstlist from 5 to
> 10 for the verlet scheme only.
>
>
> Are these results expected given the setup? Is there anything I'm doing
> wrong?
>

I am not aware of a case where compiling without OpenMP is
necessary/desirable for maximum performance. Whether you should use it (and
how many OpenMP threads to use) depends on lots of things, particularly
including what hardware you use and the atoms/core. You should experiment
if you want maximum performance.

Mark


> Thanks in advance,
> Djurre de Jong
> --
> gmx-users mailing list    gmx-users at gromacs.org
> http://lists.gromacs.org/mailman/listinfo/gmx-users
> * Please search the archive at
> http://www.gromacs.org/Support/Mailing_Lists/Search before posting!
> * Please don't post (un)subscribe requests to the list. Use the
> www interface or send it to gmx-users-request at gromacs.org.
> * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
>