Improving the Performance of Lattice Boltzmann Method with Pipelined Algorithm on A Heterogeneous Multi-zone Processor

Qingyang Zhang, Lei Xu, Rongliang Chen, Lin Chen, Xinhai Chen, Qinglin Wang, Jie Liu, Bo Yang

Abstract: Lattice Boltzmann method (LBM) has become a powerful method in computational fluid dynamics and has drawn more and more attention in high-performance computing due to its particulate nature and local dynamics, especially on recent multi-core or many-core platforms. This paper develops a parallel software framework for 3D LBM simulation on a heterogeneous multi-zone processor, MT-3000. An improved pipelined algorithm named Pencil-H is proposed, which can not only fully exploit the advantages of each component of MT-3000 but also overlap the time of calculation and communication. Moreover, an architecture-aware multi-level parallelization algorithm is developed to fully utilize the computational performance of MT-3000. A benchmark test is performed to verify the reliability and test the performance of the LBM code. Experimental results show that the optimized code achieves a 32.02\(\times \) speedup compared with using 16 CPU cores and achieves a performance of 286.03MLUPS which reaches 72.3% of the theoretical peak performance.