I greatly appreciate the flexibility and the extend of the LAMMPS package together with the detailed support provided by both manual and mailing list.
However, I was wondering if guides or tutorials exist for using LAMMPS together with CUDA?
I managed to build my LAMMPS executables for my local desktop and our HPC in a way that CUDA is supported. It provides a sufficient performance boost while running the Lennard-Jones examples in the bench-folder, however when I switch to my current systems I can hardly detect any gain in performance.
I use the following settings for cmake, which seem to work well:
cmake -D BUILD_OMP=on -D PKG_OPT=on -D PKG_USER-OMP=on -D PKG_GPU=on -D GPU_API=cuda -D PKG_KOKKOS=on -D GPU_PREC=mixed -D PKG_ASPHERE=on -D PKG_BODY=on -D PKG_COLLOID=on -D PKG_CORESHELL=on -D PKG_DIPOLE=on -D PKG_MANYBODY=on -D PKG_MC=on -D PKG_MISC=on -D PKG_MOLECULE=on -D PKG_RIGID=on -D PKG_USER-DRUDE=on -D PKG_USER-MEAMC=on -D GPU_ARCH=sm_61 …/cmake
The GPU is a NVIDIA Quadro P600.
The LAMMPS executable I built is clearly running simulations on my GPU while using the following command: mpirun –np X lmp –in in.lj –sf gpu –pk gpu 1 (i.e., using the bench Lennard-Jones file)
Therefore, I am asking if there are any general guidelines or tutorials for writing input for LAMMPS, which performs well with CUDA?
I just would like to know if there is an already existing documentation who to enhance the input for my future simulations.
Nevertheless, I am well aware that the efficiency of GPU-accelerated simulations highly depends on the properties of the investigated systems and that providing comprehensive CUDA guidelines might be complicated and cumbersome.
I would like to thank you very much for any help and references regarding CUDA-convenient creation of LAMMPS input.
Dr. Michael Bley
Applied Theoretical Physics - Computational Physics
Albert-Ludwigs-University, Freiburg im Breisgau