Acceleration of Hydro Poro-Elastic Damage Simulation in a Shared-Memory Environment1

Abstract

Hydro-PED [1] is a numerical simulation software which models nucleation and propagation of damage zones and seismicity patterns induced by wellbore fluid injection. While most of the studies in geo-physical simulation acceleration and parallelization usually focus on exascale scenarios which are translated into vast meshes, encouraging a distributed fashion of parallelization, the nature of the current simulations of Hydro-PED dictates amount of data that can conveniently fit on a single compute node – NUMA and accelerator memory alike. Thus shared-memory parallelization (such as OpenMP) can be fully implemented. In order to utilize this insight, Hydro-PED was interfaced with Trilinos [2] linear algebra solvers package, which enabled an evolution to iterative methods such as CG and GMRES. Additionally, several code sectors were parallelized and offloaded to an accelerator using OpenMP in a fine grained manner. The changes implemented in Hydro-PED gained a total speedup of x5-x12, which will enable Hydro-PED to calculate long-term simulation scenarios of hundreds of years in a feasible time – a few weeks rather than a year.