Accelerating earthquake simulations on general-purpose graphics processors

摘要

Parallelization strategies are presented for Virtual Quake, a numerical simulation code for earthquakes basedon topologically realistic systems of interacting earthquake faults. One of the demands placed upon the simulationis the accurate reproduction of the observed earthquake statistics over three to four decades. Thisrequires the use of a high-resolution fault model in computations, which demands computational power thatis well beyond the scope of off-the-shelf multi-core CPU computers. However, the recent advances ingeneral-purpose graphic processing units have the potential to address this problem at moderate cost increments.A functional decomposition of Virtual Quake is performed, and opportunities for parallelization arediscussed in this work. Computationally intensive modules are identified, and these are implemented ongraphics processing units, significantly speeding up earthquake simulations. In the current best case scenario,a computer with six graphics processing units can simulate 500 years of fault activity in Californiaat 1.5km × 1.5km element resolution in less than 1 hour, whereas a single CPU requires more than 2 daysto perform the same simulation.