[gmx-developers] Segmentation fault switching from group to verlet cutoff shceme

[David van der Spoel]

On 2013-03-16 23:24, Bu Wang wrote:
> Hi Berk,
>
> It might be rare in biological simulations, but for simulations of
> inorganic materials, it is very common. I think it is a shame that the
> scope of GROMACS is limited to biomaterials. Lots of researches in my
> field could benefit from the GROMACS' efficiency and functionality. And
> some small improvements would enable that.
>
> Anyway, I guess we will have to invest some effort to figure this out
> ourselves. To help us get started, do we need cuda programing
> experiences to modify this part of the code?

How about support for tables on GPU?

>
> Thanks,
> Bu
>
>
> On Sat, Mar 16, 2013 at 12:09 PM, Berk Hess <hess at kth.se
> <mailto:hess at kth.se>> wrote:
>
>     Hi,
>
>     Buckingham is not a high priority, as it is rarely used.
>     But you can easily edit the cuda kernel and put in the Buckingham
>     formula.
>     The main issue would be to pass the parameters, as LJ uses two
>     paramater,
>     but Buckingham three.
>
>     Cheers,
>
>     Berk
>
>     On 03/16/2013 04:21 PM, Bu Wang wrote:
>     Berk,
>
>     Thank you for clarifying the issue. Do hope that the support for
>     Buckingham potentials could be added soon so we can utilize our GPU
>     cards.
>
>     Regards,
>     Bu
>
>     On Sat, Mar 16, 2013 at 5:07 AM, Berk Hess <hess at kth.se
>     <mailto:hess at kth.se><mailto:hess at kth.se <mailto:hess at kth.se>>> wrote:
>     Hi,
>
>     Yes, that is probably the issue.
>     I forgot to add a check for Buckingham. I will add this to grompp
>     and mdrun.
>
>     Cheers,
>
>     Berk
>
>     On 03/16/2013 03:52 AM, Bu Wang wrote:
>     Hi Mark,
>
>     Thanks for your attention.
>
>     Taking the group scheme .tpr and -testverlet fails in the same way.
>     I've opened an issue as per your instruction.
>
>     I just noticed that the manual says the Verlet scheme only works
>     with plain LJ. We are using Buckingham potentials for non-bonded
>     interactions. Could this be the issue? Grompp didn't complain though.
>
>     Thanks,
>     Bu
>
>     On Fri, Mar 15, 2013 at 5:48 PM, Mark Abraham
>     <mark.j.abraham at gmail.com
>     <mailto:mark.j.abraham at gmail.com><mailto:mark.j.abraham at gmail.com
>     <mailto:mark.j.abraham at gmail.com>><mailto:mark.j.abraham at gmail.com
>     <mailto:mark.j.abraham at gmail.com><mailto:mark.j.abraham at gmail.com
>     <mailto:mark.j.abraham at gmail.com>>>> wrote:
>
>
>     On Fri, Mar 15, 2013 at 8:58 PM, Bu Wang <bw2 at alfred.edu
>     <mailto:bw2 at alfred.edu><mailto:bw2 at alfred.edu
>     <mailto:bw2 at alfred.edu>><mailto:bw2 at alfred.edu
>     <mailto:bw2 at alfred.edu><mailto:bw2 at alfred.edu
>     <mailto:bw2 at alfred.edu>>><mailto:bw2 at alfred.edu
>     <mailto:bw2 at alfred.edu><mailto:bw2 at alfred.edu
>     <mailto:bw2 at alfred.edu>><mailto:bw2 at alfred.edu
>     <mailto:bw2 at alfred.edu><mailto:bw2 at alfred.edu
>     <mailto:bw2 at alfred.edu>>>>> wrote:
>     Hello,
>
>     We encountered a problem with the verlet cutoff scheme in GROMACS
>     4.6.1. A simulation that works fine with the group scheme will crash
>     with "Segmentation fault: 11" error using the verlet method. GDB
>     shows that it happens in pme.c: 415 idxptr[ZZ] = pme->nnz[tiz], but
>     we are not sure how to proceed from here.
>
>     Sounds suspicious. Please open an issue at redmine.gromacs.org
>     <http://redmine.gromacs.org><http://redmine.gromacs.org><http://redmine.gromacs.org><http://redmine.gromacs.org>
>     with a description of your simulation, and preferably .mdp and .tpr
>     files.
>
>     It is possible that our simulation is blowing up, as we can see the
>     non-bonded energy becomes absurd in the very first step after
>     switching to verlet scheme. However, we don't see how it is possible
>     that the same simulation would work with the group scheme. The group
>     scheme gives correct energies, and can run to the end without any
>     problem.
>
>     That does make it sound like it might be a bug. Can you take your
>     group scheme .tpr and use mdrun -testverlet successfully?
>
>     To further diagnose the problem, we are wondering how we can let
>     GROMACS output the neighbor list. We would also greatly appreciate
>     any suggestion to resolve the problem.
>
>     The construction of the neighbour lists is different between group
>     and Verlet kernels, so there is not much to gain from inspecting
>     them. Knowing the combination of algorithms in use is of more
>     interest (so .mdp and any mdrun command line options).
>
>     Mark
>
>
>


-- 
David van der Spoel, Ph.D., Professor of Biology
Dept. of Cell & Molec. Biol., Uppsala University.
Box 596, 75124 Uppsala, Sweden. Phone:	+46184714205.
spoel at xray.bmc.uu.se    http://folding.bmc.uu.se