[gmx-developers] electrostatic forces and domain decomposition

[Mark Abraham]

On 27/06/2011 2:15 AM, Dommert Florian wrote:
> On Sun, 2011-06-26 at 17:44 +0200, Dommert Florian wrote:
>> On Sun, 2011-06-26 at 16:54 +0200, Dommert Florian wrote:
>>> On Sun, 2011-06-26 at 08:45 -0400, Roland Schulz wrote:
>>>> Hi,
>>>>
>>>> On Sun, Jun 26, 2011 at 8:31 AM, Dommert Florian
>>>> <dommert at icp.uni-stuttgart.de>  wrote:
>>>>          with the current release-4-5-patches branch unexpected large
>>>>          electrostatic forces arise when using PME and MPI in
>>>>          combination with domain decomposition. I tried to attach a
>>>>          small test case with a water box of about 4000 molecules and a
>>>>          corresponding mdp, top, itp, and gro file, but it was too
>>>>          large. For particle decomposition
>>>>          with different numbers of processes everything is fine as well
>>>>          as for a
>>>>          run with a single process and DD. However using multiple
>>>>          processes and
>>>>          DD I obtain forces around 10^10, though they should be around
>>>>          10^3.
>>>>
>>>>          For the testruns I compiled gromacs with cmake on a
>>>>          Ubuntu 10.4 platform with a 2.6.39-rc6+ kernel using gcc4.4.3
>>>>          and
>>>>          OpenMPI 1.4.3 and fftw3.2.2
>>>>          and a Rocks Cluster with a 2.6.18-164.6.1.el5 kernel and the
>>>>          same
>>>>          compiler, MPI and fftw3 version. On both systems I obtain the
>>>>          strange
>>>>          behaviour.
>>>>
>>>>          When I got back to the commit with SHA1 ID
>>>>          e5eb052ebf35e4b778f7a3908463a68a2ef6faee, merge by merge,
>>>>          everything is fine again.
>>>>
>>>>
>>>> What do you mean merge by merge? Could you do a git bisect?
>>> Hello,
>>>
>>> after bisecting the release-4-5-patches branches, git tracked down the
>>> error to commit:
>>>
>>> [68ab9f6fa143fec070ffe60686cb8f9e397d0376] Expanded LINCS warning text
>>>
>>> Is this what you asked me to do ?
>>>
>> Now I did it again, but ended up with another result. The output of git
>> bisect good was:
>> 2e4f7f300a284690ef2ad60359f3d7e40653724c is the first bad commit ...
>>
>> /Flo
>>
> As I tested the "oviously" working version, it turned out, that the
> results is of random behaviour. Though I turned off fftw-measurement and
> used the -reprod flag of mdrun, the average maximum force has values
> that arise randomly and are in an interval from 10^3 to 10^10. Also for
> the code I considered as "working" this behaviour can be observed.

I don't understand what version you regard as "obviously" working, or 
even working.

The commits with hashes you have mentioned can't be causing such 
problems, so if you have used git bisect correctly, then the problem is 
actually not in the code.

Are you saying that a range of versions of release-4-5-patches give you 
maximum force values up to 10^10? If so, please file a Redmine issue 
http://redmine.gromacs.org <http://redmine.gromacs.org/>, where you can 
make suitable attachments.

My suspicion is that you are experiencing some kind of linking error, 
perhaps between different available FFTW or GROMACS versions. Perhaps 
you can try static linking.

Mark

> Any idea where this comes from ?
>
> /Flo
>
>>> /Flo
>>>>
>>>> Roland
>>>>
>>>>
>>>>          /Flo
>>>>          --
>>>>          Florian Dommert
>>>>          Dipl. - Phys.
>>>>
>>>>          Institute for Computational Physics
>>>>          University Stuttgart
>>>>
>>>>          Pfaffenwaldring 27
>>>>          70569 Stuttgart
>>>>
>>>>          EMail: dommert at icp.uni-stuttgart.de
>>>>          Homepage: http://www.icp.uni-stuttgart.de/~icp/Florian_Dommert
>>>>
>>>>          Tel.: +49 - (0)711 - 68563613
>>>>          Fax.: +49 - (0)711 - 68563658
>>>>
>>>>
>>>>
>>>>
>>>> -- 
>>>> ORNL/UT Center for Molecular Biophysics cmb.ornl.gov
>>>> 865-241-1537, ORNL PO BOX 2008 MS6309
>>>>
>>>> -- 
>>>> gmx-developers mailing list
>>>> gmx-developers at gromacs.org
>>>> http://lists.gromacs.org/mailman/listinfo/gmx-developers
>>>> Please don't post (un)subscribe requests to the list. Use the
>>>> www interface or send it to gmx-developers-request at gromacs.org.
>>>
>>> -- 
>>> gmx-developers mailing list
>>> gmx-developers at gromacs.org
>>> http://lists.gromacs.org/mailman/listinfo/gmx-developers
>>> Please don't post (un)subscribe requests to the list. Use the
>>> www interface or send it to gmx-developers-request at gromacs.org.
>>
>> -- 
>> gmx-developers mailing list
>> gmx-developers at gromacs.org
>> http://lists.gromacs.org/mailman/listinfo/gmx-developers
>> Please don't post (un)subscribe requests to the list. Use the
>> www interface or send it to gmx-developers-request at gromacs.org.
>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://maillist.sys.kth.se/pipermail/gromacs.org_gmx-developers/attachments/20110627/0fa69038/attachment.html>