Weak scaling performance, Sphere benchmark, K80, 1M particles/node
Performance in millions of particle-timesteps / second / node
Nodes | Kokkos/Cuda-1 (mpi) | Kokkos/Cuda-2 (mpi) | |
1 | 182.7 (2) | 252.6 (4) | |
2 | 170.5 (2) | 225.8 (4) | |
4 | 150.9 (2) | 196.3 (4) | |
8 | 123.2 (2) | 161.2 (4) |
1 | mpirun -np 2 --npersocket 1 --bind-to core spa_ride80_kokkos_cuda -sf kk -k on g 2 -pk kokkos reduction atomic comm threaded -v x 64 -v y 40 -v z 40 -v t 100 -in in.sphere.gpu.steps -log log.sparta.date=23Dec17.model=sphere.machine=ride80.pkg=kokkos_cuda.kind=weak.size=1M.node=1.mpi=2.gpu=2 |
2 | mpirun -np 4 --npersocket 1 --bind-to core spa_ride80_kokkos_cuda -sf kk -k on g 2 -pk kokkos reduction atomic comm threaded -v x 64 -v y 40 -v z 80 -v t 100 -in in.sphere.gpu.steps -log log.sparta.date=23Dec17.model=sphere.machine=ride80.pkg=kokkos_cuda.kind=weak.size=1M.node=2.mpi=2.gpu=2 |
4 | mpirun -np 8 --npersocket 1 --bind-to core spa_ride80_kokkos_cuda -sf kk -k on g 2 -pk kokkos reduction atomic comm threaded -v x 64 -v y 80 -v z 80 -v t 100 -in in.sphere.gpu.steps -log log.sparta.date=23Dec17.model=sphere.machine=ride80.pkg=kokkos_cuda.kind=weak.size=1M.node=4.mpi=2.gpu=2 |
8 | mpirun -np 16 --npersocket 1 --bind-to core spa_ride80_kokkos_cuda -sf kk -k on g 2 -pk kokkos reduction atomic comm threaded -v x 128 -v y 80 -v z 80 -v t 100 -in in.sphere.gpu.steps -log log.sparta.date=23Dec17.model=sphere.machine=ride80.pkg=kokkos_cuda.kind=weak.size=1M.node=8.mpi=2.gpu=2 |
1 | mpirun -np 4 --npersocket 2 --bind-to core spa_ride80_kokkos_cuda -sf kk -k on g 4 -pk kokkos reduction atomic comm threaded -v x 64 -v y 40 -v z 40 -v t 100 -in in.sphere.gpu.steps -log log.sparta.date=23Dec17.model=sphere.machine=ride80.pkg=kokkos_cuda.kind=weak.size=1M.node=1.mpi=4.gpu=4 |
2 | mpirun -np 8 --npersocket 2 --bind-to core spa_ride80_kokkos_cuda -sf kk -k on g 4 -pk kokkos reduction atomic comm threaded -v x 64 -v y 40 -v z 80 -v t 100 -in in.sphere.gpu.steps -log log.sparta.date=23Dec17.model=sphere.machine=ride80.pkg=kokkos_cuda.kind=weak.size=1M.node=2.mpi=4.gpu=4 |
4 | mpirun -np 16 --npersocket 2 --bind-to core spa_ride80_kokkos_cuda -sf kk -k on g 4 -pk kokkos reduction atomic comm threaded -v x 64 -v y 80 -v z 80 -v t 100 -in in.sphere.gpu.steps -log log.sparta.date=23Dec17.model=sphere.machine=ride80.pkg=kokkos_cuda.kind=weak.size=1M.node=4.mpi=4.gpu=4 |
8 | mpirun -np 32 --npersocket 2 --bind-to core spa_ride80_kokkos_cuda -sf kk -k on g 4 -pk kokkos reduction atomic comm threaded -v x 128 -v y 80 -v z 80 -v t 100 -in in.sphere.gpu.steps -log log.sparta.date=23Dec17.model=sphere.machine=ride80.pkg=kokkos_cuda.kind=weak.size=1M.node=8.mpi=4.gpu=4 |