Autotuning of Adaptive Mesh Refinement PDE Solvers on Shared Memory Architectures

Abstract

Many multithreaded, grid-based, dynamically adaptive solvers for partial differential equations permanently have to traverse subgrids (patches) of different and changing sizes. The parallel efficiency of this traversal depends on the interplay of the patch size, the architecture used, the operations triggered throughout the traversal, and the grain size, i.e. the size of the subtasks the patch is broken into. We propose an oracle mechanism delivering grain sizes on-the-fly. It takes historical runtime measurements for different patch and grain sizes as well as the traverse's operations into account, and it yields reasonable speedups. Neither magic configuration settings nor an expensive pre-tuning phase are necessary. It is an autotuning approach. © 2012 Springer-Verlag.