GPU-accelerated OpenFOAM Simulations with PETSc

This technical report presents a preliminary comparison of OpenFOAM's embedded linear algebra solvers with external solvers from the PETSc library and other packages, focusing on GPU support. The objective is to evaluate the performance of standard solvers on CPUs versus algebraic multi-grid based solvers on hybrid CPU-GPU systems. The tests were conducted on two architectures: Intel Skylake and IBM Power9 with Nvidia V100 GPUs. Preliminary results indicate a performance gain of approximately 5x in serial execution when comparing 1 CPU to 1 CPU + 1 GPU, while the full node configuration showed a speedup of about 2x. The report details the methodology, including the software stack and test-case used, which involved a 3-D Lid-driven cavity flow simulation. Performance results are documented, highlighting the influence of the BLAS library version on overall performance. The findings serve as a foundation for further work in optimizing solver configurations for enhanced computational efficiency.