Dear team,
I have run two versions of RapidCFD on an HPC cluster with 2 NVIDIA GPUs per node:
RapidCFD A: OpenMPI-3.0.0(hwloc enabled) + CUDA-10.0
RapidCFD B: OpenMPI-5.0.8(internal hwloc enabled) + CUDA-12.0 + UCX + gdrcopy + nv_peer_mem
The CFD model used was a 3D Lid-Driven Cavity (8.6M cells).https://develop.openfoam.com/committees/hpc/-/tree/develop/incompressible/icoFoam/old?ref_type=heads
Observations:
Single Node (2 GPUs): RapidCFD B is not performing better than RapidCFD A.
Two Nodes (4 GPUs): RapidCFD B performs slightly better than RapidCFD A.
I am looking for ways to improve the single-node performance of RapidCFD B to at least match or exceed RapidCFD A. I would highly appreciate any suggestions or input on optimization
Dear team,
I have run two versions of RapidCFD on an HPC cluster with 2 NVIDIA GPUs per node:
RapidCFD A: OpenMPI-3.0.0(hwloc enabled) + CUDA-10.0
RapidCFD B: OpenMPI-5.0.8(internal hwloc enabled) + CUDA-12.0 + UCX + gdrcopy + nv_peer_mem
The CFD model used was a 3D Lid-Driven Cavity (8.6M cells).https://develop.openfoam.com/committees/hpc/-/tree/develop/incompressible/icoFoam/old?ref_type=heads
Observations:
Single Node (2 GPUs): RapidCFD B is not performing better than RapidCFD A.
Two Nodes (4 GPUs): RapidCFD B performs slightly better than RapidCFD A.
I am looking for ways to improve the single-node performance of RapidCFD B to at least match or exceed RapidCFD A. I would highly appreciate any suggestions or input on optimization