Abstract

We present a memory- and cache-efficient approach for simulations on recursively refined dynamically adaptive triangular grids. Grid cells are stored and processed in an order defined by the Sierpinski curve; the resulting locality properties are exploited for optimised serial implementation and parallelisation. The approach is particularly designed for Finite Volume and discontinuous Galerkin solvers, with Tsunami simulation as the main target application. In the talk, we will discuss approaches for parallelisation in shared and distributed memory. We will present a classical partitioning-based strategy, as well as a novel shared-memory approach based on the dynamical scheduling of many small sub-partitions. Here, the intention is to allow for strongly varying computational load per element (as required for inundation modelling, or for local time-stepping methods).

In addition, we would like to discuss some ideas on how to provide bathymetry data and (time-dependent) displacements for a simulation with dynamically adaptive refinement. We’ll present some first results of adaptive simulations using the augmented Riemann solvers provided with GeoClaw.