Hi all. Like the title, I can’t enable the KOKKOS package. I’ve used OpenMP with no issues. I am simulating heat transfer using the Green-Kubo method in stacked boron nitride/graphene sheets. I want to use KOKKOS because the hybrid/overlay command only has the /kk accelerator, without it the simulation is exceedingly long. I’m open to changing pair styles as well if anyone has a suggestion.
I have the commands “package omp 16”, “package kokkos”, “suffix hybrid omp kk” in the input script. To run the script I input “lmp -in BNGR-GK.in -k on” into the command line which results in the following
It’s a warning, meaning the script will run fine as is, although the environment variables listed there are suggested for improved performance. See 6 Environment Variables for detailed documentation on OpenMP environment variables and how to set them.
Hybrid pair styles do not need to have accelerator support except for KOKKOS (due to the different memory handling in KOKKOS. So your basic conclusion here is wrong.
Since you are using Windows, why don’t you just install an MPI version of LAMMPS and use MPI parallelization? This is usually more efficient that multi-threading and does not require any accelerator support at all.
Since you are not providing any details about your input, it is practically impossible to make any suggestions.
This is a VERY bad idea. You should not mix KOKKOS with any other accelerator. So you either must go “all-KOKKOS” or “no-KOKKOS”.
I would need to see the actual input script and also some timing summary output to document that that actual source of slow computation can be addressed with using accelerated pair styles.
If it exits without any output it could be encountering a segmentation fault. Like Axel said, if you can provide a minimal working example we can take a look.
Hi, thanks for the input. I’m very new to parallelization so please forgive some of the rookie mistakes
Since you are using Windows, why don’t you just install an MPI version of LAMMPS and use MPI parallelization? This is usually more efficient that multi-threading and does not require any accelerator support at all.
I’ve just done this and I’ve gotten a very obvious increase in performance
This is a VERY bad idea. You should not mix KOKKOS with any other accelerator. So you either must go “all-KOKKOS” or “no-KOKKOS”.
Yes, I figured as much after some testing. I’ve reconfigured the input to only use kokkos. I also am running the script with the mpiexec command, which I think is the correct method.
I would need to see the actual input script and also some timing summary output to document that that actual source of slow computation can be addressed with using accelerated pair styles.
The log file is taken mid script-run as I have just gotten everything to run. It does have some timing outputs for the initial equilibration. The command used to run the script is “mpiexec -np 7 lmp -k on t 2 -in BNGR-GK.in” and for hardware, I am running a single CPU Windows PC with an AMD R7 5800H (8C16T). I’ve tried GPU acceleration with my RX6600M but the driver keeps crashing and aborting the simulation, which I don’t think is a LAMMPS issue
