Problems encountered when using Kokkos in lammps for GPU acceleration

LAMMPS LAMMPS Installation
,
1

I want to use the reaxff force field in lamps to perform molecular dynamics simulations. My system is relatively simple, which is the reaction of toluene, oxygen and carbon monoxide at high temperatures. However, this reaction is very slow when using the CPU to perform parallel calculations, so I thought about using kokkos and cuda for gpu acceleration. My in file is as follows:

units	real  ##fs  Å
boundary 	p p p
atom_style	full

read_data	10_0_11_0_11.data
pair_style	reaxff NULL safezone 4.0 mincap 100
pair_coeff	* * CHO.ff C H O
fix 1 all qeq/reaxff 1 0.0 10.0 1.0e-6 reaxff


neighbor	2.0 bin
neigh_modify	every 1 delay 0 check yes


minimize    1e-10 1e-10 10000 10000

fix     11      all nve
fix     22      all temp/berendsen 2500 2500 500
fix	5 	all reaxff/species 1 100 100 species.out element C H O
fix	6 	all reaxff/bonds 1000 bonds.reaxc

timestep	0.10
thermo_style	custom  step time temp  etotal  pe ke density press 
thermo	100

dump		2 all custom 500 mydump-nvt.lammpstrj id mol type x y z  q 
dump		xtcout all  xtc 100 myxtc-nvt.xtc

run		100000000
restart         1000000 restart.equil
write_restart 	restart_.equil1

The command to use is:

lmp -in in.NVT -k on g 1 -sf kk -pk kokkos neigh half newton on

At this time, the calculation can still run normally, but it prompts me that I cannot use kokkos for acceleration due to fix-related issues. The specific error reported is as follows:

WARNING: Fix with atom-based arrays not compatible with sending data in Kokkos communication, switching to classic exchange/border communication (src/KOKKOS/comm_kokkos.cpp:666)
WARNING: Fix with atom-based arrays not compatible with Kokkos sorting on device, switching to classic host sorting (src/KOKKOS/atom_kokkos.cpp:178)

This gives me a headache, because these fix commands are the core steps of my calculations. I hope you can give me some suggestions to help me use kokkos normally.

2

Please always report which LAMMPS version you are using, what platform you are running on, and hardware you have.

These are not errors but warnings and they don’t indicate that you cannot use KOKKOS but rather that it won’t run at maximum efficiency.

The important question here is: do your calculations run faster than on the CPU?

3

First of all, I apologize for my incomplete expression before. The additional information is as follows:
My system is ubuntu22.04;
My lamps is lamps-2Aug2023, and the kokkos in it was replaced with version 4.1;
My graphics card is Titan V.
The in file I showed above is the file I used before when using pure CPU calculations.
My problem is mainly that I encountered problems when using kokkos for gpu acceleration using the previous in file, and could not correctly enable the acceleration effect of kokkos. In fact, the running result is relatively slow.
I want to see if I can correctly use kokkos for gpu acceleration by modifying the in file to obtain higher performance than pure CPU.
Thanks for your answer.

4

There is no Kokkos version of fix temp/berendsen so that will slow down the simulation. Can you use a different thermostat like fix langevin or fix nvt which has Kokkos support?

Also, how many atoms are in the system? Typically you need hundreds of thousands to a million (more is better) per GPU to start saturating the GPU.

5

I ported fix temp/berendsen to Kokkos, it is now merged into the develop branch of LAMMPS on Github.