Evaluating Performance Portability of SYCL and Kokkos: A Case Study on LBM Simulations

Since modern high performance computing systems are evolving towards diverse and heterogeneous architectures, the emergence of high-level portable programming models leads to a particular focus on performance portability. In this paper, we evaluate the performance portability and explore performance optimization methods for two portable programming models SYCL and Kokkos. We take an open-source multi-phase Lattice Boltzmann Method (LBM) flow simulation code as a case study and implement portable versions with different optimizations. Then we compare our portable implementations with engineer-tuned OpenMP and CUDA versions on Intel CPUs and NVIDIA GPUs. Experimental results show that both SYCL and Kokkos can deliver superior performance than traditional programming models, but the best performance of the portable versions depends heavily on platform-specific optimizations. There is no single implementation that can achieve the best performance on both CPUs and GPUs. Consequently, we conclude that the performance portability still needs to be further improved for both SYCL and Kokkos. In addition, we present a comparative analysis of different optimization methods that qualify the performance enhancement when using SYCL and Kokkos on CPUs and GPUs. Our work offers valuable references for the development of both portable programming models and applications.