[gmx-users] Different energies of GPU and CPU-only runs with Gromacs 5.x

[Machtens, Jan-Philipp]

Thanks for the reply!
However, the differences between GPU runs and CPU-only runs for
Gromacs 5.x that I observe are statistically significant (10-100 ns for each test sim) (and reproducible).
In contrast, with Gromacs 4.6.x I do not have significant differences that
I can apparently attribute to the use of GPUs.
Is there any other reason for the deviation with GPU and GROMAC 5.x ?
Best,
JP
________________________________________
From: gromacs.org_gmx-users-bounces at maillist.sys.kth.se [gromacs.org_gmx-users-bounces at maillist.sys.kth.se] on behalf of Mark Abraham [mark.j.abraham at gmail.com]
Sent: Monday, September 22, 2014 10:25 AM
To: Discussion list for GROMACS users
Subject: Re: [gmx-users] Different energies of GPU and CPU-only runs with Gromacs 5.x

On Mon, Sep 22, 2014 at 9:44 AM, Machtens, Jan-Philipp <
j.machtens at fz-juelich.de> wrote:

> Dear all,
> if I compare GPU-accelerated and CPU-only runs, I get significantly
> different results/energies with
> Gromacs 5.x. In contrast, with Gromacs 4.x everything is fine, i.e. no
> difference between CPU-only runs and with GPU.
>
> In short, with Gromacs 5.x using GPUs, I have lower total energy, higher
> short-range LJ, higher short-range Coulomb and
> lower reciprocal Coulomb values.


This is a normal consequence of different PME load balancing between CPU
and GPU. You can use the UNIX diff tool to observe that mdrun is making
different decisions.

This also leads to altered density and box dimensions, etc....
>

MD trajectories are generally not reproducible even if the PME load is
balanced the same way. Do replicates show the same kind of variation?

Mark

The simulations were run on a single node with either 2x GeForce GTX 780 Ti
> or with 1x GTX 770.
> Compilers tested: gcc 4.8 and icc 14
> Cuda Versions: 6.0 or 6.5
> ForceField: Amber99sb-ildn
>
> The topol.tpr was identical for each simulation. Verlet cut-off scheme was
> used.
> The mdp file can be accessed here:
> https://www.dropbox.com/s/kjkglu2p8p11ajn/parameter.mdp?dl=0
> Regression tests of the various compilations were always passed.
>
> Does someone has the some problems or has a clue what might cause the
> problems I have with Gromacs5.x when using GPUs ?
> Many thanks!
> Best,
> Jan-Philipp Machtens
> Postdoctoral Research Fellow
> Institute of Complex Systems - Cellular Biophysics (ICS-4)
> Forschungszentrum Jülich, Germany
>
>
>
> ------------------------------------------------------------------------------------------------
>
> ------------------------------------------------------------------------------------------------
> Forschungszentrum Juelich GmbH
> 52425 Juelich
> Sitz der Gesellschaft: Juelich
> Eingetragen im Handelsregister des Amtsgerichts Dueren Nr. HR B 3498
> Vorsitzender des Aufsichtsrats: MinDir Dr. Karl Eugen Huthmacher
> Geschaeftsfuehrung: Prof. Dr.-Ing. Wolfgang Marquardt (Vorsitzender),
> Karsten Beneke (stellv. Vorsitzender), Prof. Dr.-Ing. Harald Bolt,
> Prof. Dr. Sebastian M. Schmidt
>
> ------------------------------------------------------------------------------------------------
>
> ------------------------------------------------------------------------------------------------
>
> --
> Gromacs Users mailing list
>
> * Please search the archive at
> http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before
> posting!
>
> * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
>
> * For (un)subscribe requests visit
> https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users or
> send a mail to gmx-users-request at gromacs.org.
>
--
Gromacs Users mailing list

* Please search the archive at http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before posting!

* Can't post? Read http://www.gromacs.org/Support/Mailing_Lists

* For (un)subscribe requests visit
https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users or send a mail to gmx-users-request at gromacs.org.