Performance analysis of a parallel implementation of the lattice Boltzmann method for computational fluid dynamics.

摘要

We propose to use benchmark data combined with detailed O(n) analysis to predict the performance of a parallel Lattice-Boltzmann Method (LBM) for 2D fluid dynamics simulation with solid particles on various configurations of cluster computers. The LBM has super step synchronism, phase concurrent components and non-critical size of division properties. Our results demonstrate accurate predictions for LBM simulation performance. The CPU benchmark indicates that increased variable data precision does not degrade execution time significantly on Pentium class processors. We show that improved communication and calculation strategies for solid particles yielded better speedup and scalability. A theoretical analysis demonstrates that worst case speedup occurs when all the solid particles saturate a single workspace.