Performance analysis of the graph-partitioning algorithms used in OpenFOAM

Abstract

OpenFOAM is a widely used opensource CFD application. Based on mesh partitioned, applications can run in parallel to achieve better performance in OpenFOAM. When mesh generated from the liquid field is large, performance of partitioning algorithms will heavily affect the execution efficiency of the whole application. In this paper, we investigate the four partitioning algorithms implemented in OpenFOAM-Simple, Hierarchical, Scotch and Metis and analyze their performance. Performance evaluation includes partitioning time, communication overhead, quality of load balancing and application's parallel execution time based on the experiment of LinearPTT with 34,800,000 cells on Tianhe-1A. The results show that Scotch spends the most time on partitioning and Metis' partitioning time remains steady when the number of processors increases. Both Scotch and Metis introduce less communication overhead than Simple and Hierarchical. Scotch does better in balancing the cells among processors than Metis. When scales of meshes and numbers of processors increase, only Scotch and Metis are practicable with acceptable performance. But their partitioning efficiency still needs a betterment.