## Poster 97: Optimizing Multigrid Poisson Solver of Cartesian CFD Code CUBE

**Authors:** Kazuto Ando (RIKEN Center for Computational Science (R-CCS)), Rahul Bale (RIKEN), Keiji Onishi (RIKEN Center for Computational Science (R-CCS)), Kiyoshi Kumahata (RIKEN Center for Computational Science (R-CCS)), Kazuo Minami (RIKEN Center for Computational Science (R-CCS)), Makoto Tsubokura (Kobe University, RIKEN Center for Computational Science (R-CCS))

**Abstract:** We demonstrate an optimization of multigrid Poisson solver of Cartesian CFD code “CUBE (Complex Unified Building cubE method)”. CUBE is a simulation framework for complex industrial flow problem, such as aerodynamics of vehicles, based on hierarchical Cartesian mesh. In incompressible CFD simulation, solving pressure Poisson equation is the most time-consuming part. In this study, we use a cavity flow simulation as a benchmark problem. With this problem, multigrid Poisson solver dominates 91% of execution time of the time-step loop. Specifically, we evaluate the performance of Gauss-Seidel loop as a computational kernel based on “Byte per Flop” approach. With optimization of the kernel, we achieved 9.8x speedup and peak floating point performance ratio increased from 0.4% to 4.0%. We also measured parallel performance up to 8,192 nodes (65,536 cores) on the K computer. With optimization of the parallel performance, we achieved 2.9x–3.9x sustainable speedup in the time-step loop.

**Best Poster Finalist (BP):** no

**Poster:** PDF

**Poster summary:** PDF

Back to Poster Archive Listing