SC13 Denver, CO

The International Conference for High Performance Computing, Networking, Storage and Analysis

Optimized Kernels for Large Scale Earthquake Simulations with SeisSol, an Unstructured ADER-DG Code.

Authors: Alexander Heinecke (Technical University of Munich), Alexander Breuer (Technical University of Munich), Sebastian Rettenberger (Technical University of Munich), Michael Bader (Technical University of Munich), Stefan Wenk (Ludwig-Maximilians-Universität-München), Alice Gabriel (Ludwig-Maximilians-Universität München), Christian Pelties (Ludwig-Maximilians-Universität-München)

Abstract: The software package SeisSol is one of the leading codes for earthquake scenarios, in particular for simulating dynamic rupture processes and for problems that require discretization of very complex geometries. SeisSol uses the discontinuous Galerkin (DG) method for spatial and Arbitrary high order DERivatives (ADER) for time discretization.We present optimizations for the inner, cell-local kernel routines of SeisSol, where SeisSol spends most of the total computing time. The kernels are implemented as a sequence of matrix-matrix-multiplications of a relatively small size and varying sparsity patterns (ranging from "very" sparse to dense). When running the LOH.1 benchmark, employing a 6th order space-time-discretization, in a strong scaling setting on up to 32,000 cores of SuperMUC, we achieved a maximum performance of 153 TFLOPS, which corresponds to 21.6% of the theoretical peak performance. By comparing our optimized version to the classic version of SeisSol we obtained a total application speed-up of roughly 4.5.

Poster: pdf
Two-page extended abstract: pdf

Poster Index