Parallel Multigrid with Adaptive Multilevel hCGA on Manycore Clusters
Performance Analysis and Optimization
TimeWednesday, June 19th2:15pm - 2:45pm CEST
DescriptionA multigrid is a scalable multilevel method for solving linear equations and preconditioning Krylov iterative linear solvers, and is especially suitable for large-scale problems because of its scalable feature. The parallel multigrid method is expected to be one of the most powerful tools on exa-scale systems. In the previous work (K. Nakajima, IEEE ICPADS 2014), we have already developed an FVM code for 3D Groundwater Flow through Heterogenous Porous Media (pGW3D-FVM) with MGCG solvers (Multigrid Preconditioned Conjugate Gradient) using OpenMP/MPI with RCM (Reverse Cuthill-Mckee), and it is ready for exa-scale systems by hCGA (Hierarchical Coarse Grid Aggregation). hCGA provided significant improvement of performance (60% in weak scaling, 600% in strong scaling) for 3D FVM application on 4,096 nodes of Oakleaf-FX for problems with 1.8×10^10 DOF. Because the hCGA can only handle 2-hierarchica-levels, we are developing AM-hCGA (Adaptive Multilevel hCGA) for multiple hierarchical levels (more than three). In this presentation, we will present preliminary results of AM-hCGA on the Oakforest-PACS, Joint Center for Advanced High Performance Computing (JCAHPC), which consists of 8,208 nodes of Intel Xeon Phi (Knights Landing).
Professor Supercomputing Research Division,