Should I always use Kokkos versions of fix commands with pair_style reaxff/kk?

LAMMPS LAMMPS General Discussion
1

Hi all,

I’m running simulations using the Kokkos-accelerated ReaxFF force field in LAMMPS, such as pair_style reaxff/kk or reaxff/kk/device. I noticed that there are also Kokkos versions of other related commands, for example:

  • fix qeq/reaxff/kk, fix qeq/reaxff/kk/device, fix qeq/reaxff/kk/host
  • fix npt/kk, fix nvt/kk, fix npt/kk/device, etc.

My question is:

When I use reaxff/kk, is it necessary (or at least strongly recommended) to also use Kokkos variants of these fixes, such as fix npt/kk and fix qeq/reaxff/kk, instead of their non-Kokkos counterparts?

Any suggestions or best practices would be greatly appreciated.

Thanks in advance!

Best regards,

2

Both approaches should yield the same results, but if you use GPU acceleration, it is more than recommended, because even a single non-KOKKOS fix can slow down the simulation significantly due to additional communication. You should use -sf kk switch when invoking LAMMPS, e.g. lmp -in input.in -sf kk, which will use KOKKOS implementations wherever it is possible.
I don’t know if it is the same for CPU version, but you can test it easily by running the simulations for both cases.

1 Like
3

Thank you very much for the detailed explanation! That really helps clarify things.

I’ll make sure to use the -sf kk flag when running KOKKOS-accelerated simulations, and I’ll also run some tests to compare performance on CPU-only runs.

Thanks again for your support!

Best regards,