Simple direct n-body solvers using Kokkos for parallelization and performance
git clone --recursive https://github.com/markstock/nvortexKokkos.git
mkdir build
cd build
ccmake ..
make
./nvKok01.bin -n=100000
The first code, nvKok01.cpp is designed to be the minimal n-body code for Kokkos. When compiled with the OPENMP driver, the performance should be similar to the basic "host" OpenMP operation.
Future versions will be designed to make calls to accelerators (via HIP or CUDA backends).
Launch a CPU-only batch job with Slurm on a Bard Peak (64-core Epyc) with these commands:
OMP_PROC_BIND=spread OMP_PLACES=threads srun -N1 -n1 -pbardpeak --exclusive --threads-per-core=1 --cpus-per-task=64 -t0:00:30 ./nvKok01.bin -n=300000
OMP_PROC_BIND=spread OMP_PLACES=threads srun -N1 -n1 -pbardpeak --exclusive --threads-per-core=2 --cpus-per-task=128 -t0:00:30 ./nvKok01.bin -n=300000