99. A Fast Implicit Solver with Low Memory Footprint and High Scalability for Comprehensive Earthquake Simulation System
Authors: Kohei Fujita (RIKEN)Tsuyoshi Ichimura (University of Tokyo)Kentaro Koyama (Fujitsu Ltd)Masashi Horikoshi (Intel Corporation)Hikaru Inoue (Fujitsu Ltd)Larry Meadows (Intel Corporation)Seizo Tanaka (University of Tsukuba)Muneo Hori (University of Tokyo)Lalith Maddegeddara (University of Tokyo)Takane Hori (Japan Agency for Marine-Earth Science and Technology)
Best Poster Finalist
Abstract: We developed a comprehensive earthquake simulation system that improves the reliability of conventional earthquake disaster estimates by significant speedup of unstructured implicit finite-element simulations needed to solve nonlinear wave-propagation problems in complex-shaped domains. A fast implicit solver with a low-memory footprint was developed using an algorithm combining a multi-grid, mixed-precision, communication avoiding inexact LU preconditioning, and element-by-element method with novel SIMD-buffering and multi-core coloring methods. Using this solver with 88.8% size-up scalability up to 663,552 CPU cores of the full K computer, a challenging practical 2.05 Tera degrees-of-freedom problem (205 times more degrees-of-freedom compared to the current state-of-the-art) was solved 30 times faster compared with a memory-efficient SC14 solver. This poster presents the algorithmic design, implementation, performance, and portability of this fast and memory-efficient implicit solver.
Two-page extended abstract: pdf