Optimization of Parallel Mesh Generation via File for Multigrid Method
Toshihiro Hanawa, Kengo Nakajima · 2019
We consider about the optimization of parallel mesh generation process on the Oakforest-PACS system (OFP) which is the largescale cluster using Intel Xeon Phi and OmniPath Architecture. OFP has not only the Lustre file system but also Infinite Memory Engine (IME) as the fast file cache system. MPI-IO, which is the parallel programming interface for file IO, provides the single shared file access, the optimal file access on the Lustre file system, and the native access to IME. As a result, we obtained 243 times better performance in maximum than the original version with the File Per Process. Moreover, the case of the Single Shared File with IME is 31 times faster than the original one with the File Per Process.