BKS parametrization of buck/coul/long pair_style for silica

Hi everyone!

  This topic also interests me. I am currently doing simulations with pair
buck/coul/long and also having some problems with the singularity at r=0.
However, my system is lithium disilicate instead of silica. Is there a
systematic way in lammps to prevent this undesired behaviour for buck/coul
and born/coul ? Is there a hybrid/overlay option for gpu accelerated
packages?

  Another question is about the performance comparison between packages
GPU and USER-CUDA. Usually the latter performs significantly better (about
20%). However, there were cases in which my simulations with USER-CUDA
yielded strange results like superdiffusivity or 'lost atoms'. I know that
these problems can arise due to the singularity of the potential. The
thing is that they are absent in simulations using CPU or the GPU package.
What could be the cause of those discrepancies?

  All the above observations are essentially the same independently of the
hardware used.

  Thanks in advance! Any help would be greatly appreciated!

  Best,
  Luis

Dear Axel,
Thank you for your suggestion.
I have previously successfully run the simulation with the gpu library

provided with LAMMPS, using pair_style table/gpu, however, I was 1)
seeking to understand the "fusion problem," 2) hoping to use the
USER-CUDA library as it seems to perform better in our installation.

with respect to point 2)
what is your specific setup and how do you run LAMMPS on it?
it would be extremely helpful for us to know more about when
either of the two GPU accelerated package performs better.
have you been running with one, two or more MPI tasks per GPU?
what kind of GPU and host hardware do you have?

thanks,
    axel.

Hi

could you send me input and data files which shows the superdiffusivity issue? I'd like to look into it. Did you do single precision simulations? I mean it could be that this is a issue coming from single precision integration - which would be really important to know.

Cheers
Christian

-------- Original-Nachricht --------