[gmx-users] Gromacs on CPU runs fine, same calculation crashes on GPU with nonbonded interaction larger than table limit

[Marc Hoemberger]

Hi Justing,

I followed your advice and  compiled GROMACS 2016.4 version and used this
vanilla version (no Plumed patch applied).

I see a very similar behavior as described beofre, in that when starting
the simulation on a CPU only node, the system behaves well (simulation is
now running for 2ns+ without any problems),  however, when I start it on
the GPU/CPU the system seems to be exploding with many LINCS warnings:

Step 231, time 0.462 (ps)  LINCS WARNING
relative constraint deviation after LINCS:
rms 172.973862, max 7198.149414 (between atoms 2104 and 2105)
bonds that rotated more than 30 degrees:
atom 1 atom 2  angle  previous, current, constraint length
2098   2101   55.5    0.1528   0.0888      0.1530
2101   2102  132.5    0.1304 796.0084      0.1111

And many more of this kind. The system has been equilibrated well
(NVT+NPT+unconstrained 10ns run - all on CPU only). I tried different
combinations of GPU/Threads, but the system just seems to not behave well,
as soon as I try to run it on the GPU node. I am using CUDA version 7.5 and
NVIDIA GeForce GTX 780 TI cards.

I uploaded both log file and error file on pastebin, in case this might
help:
LOGFILE: https://pastebin.com/q24WZ6vF
ERRORLOG: https://pastebin.com/R8aMHmnD

-Marc


> Hi everyone,
>
> I ran a targeted md simulation with gromacs 5.1.4 + plumed 2.4.0 on a
> 32-core node. The simulation worked well, however when I wanted to try to
> run the identical simulation (same starting points etc) on a CPU/GPU node
> the run crashes with two warnings of the following type:
>
> WARNING: Listed nonbonded interaction between particles X and Y at distance
> Z which is larger than the table limit W nm.
>
> I know that my system is equilibrated well, and the same setup does work
> perfect on the CPU only run.
>
> For running the GPU run I use -ntomp 4 and one gpu, with -pin on. I use
> CUDA 7.0.28 and gromacs 5.1.4 with plumed 2.4.0.
>
> If there is any additional information that could help, let me know and I
> will provide it.
>
> Best,
> Marc
>
>
> Message: 4
> Date: Wed, 28 Feb 2018 13:14:49 -0500
> From: Justin Lemkul <jalemkul at vt.edu>
> To: gmx-users at gromacs.org
> Subject: Re: [gmx-users] Gromacs on CPU runs fine, same calculation
>         crashes on GPU with nonbonded interaction larger than table limit
> Message-ID: <4e9deb70-0d37-09e7-36ff-99bf880adb5b at vt.edu>
> Content-Type: text/plain; charset=utf-8; format=flowed
>
>
>
> On 2/28/18 1:13 PM, Marc Hoemberger wrote:
> > Hi everyone,
> >
> > I ran a targeted md simulation with gromacs 5.1.4 + plumed 2.4.0 on a
> > 32-core node. The simulation worked well, however when I wanted to try to
> > run the identical simulation (same starting points etc) on a CPU/GPU node
> > the run crashes with two warnings of the following type:
> >
> > WARNING: Listed nonbonded interaction between particles X and Y at
> distance
> > Z which is larger than the table limit W nm.
> >
> > I know that my system is equilibrated well, and the same setup does work
> > perfect on the CPU only run.
> >
> > For running the GPU run I use -ntomp 4 and one gpu, with -pin on. I use
> > CUDA 7.0.28 and gromacs 5.1.4 with plumed 2.4.0.
> >
> > If there is any additional information that could help, let me know and I
> > will provide it.
>
> Can you reproduce the problem with a vanilla GROMACS version
> (non-PLUMED), preferably something newer?
>
> -Justin
>
> --
> ==================================================
>
> Justin A. Lemkul, Ph.D.
> Assistant Professor
> Virginia Tech Department of Biochemistry
>
> 303 Engel Hall
> 340 West Campus Dr.
> Blacksburg, VA 24061
>
> jalemkul at vt.edu | (540) 231-3129
> http://www.biochem.vt.edu/people/faculty/JustinLemkul.html
>
>