Optimization strategies for the acoustic wave stencil on multicore systems

Abstract

Seismic imaging techniques are extensively used in geophysical exploration. For example, the full-waveform inversion (FWI) and the reverse time migration (RTM) are essential applications for the identification and placement of hydrocarbon reservoirs and the characterization of the subsurface material like porosity, viscosity, acoustic velocity, localization, dimensions, and others. Such applications typically rely on finite differences methods and stencil computation kernels to solve the physics of interest, which demand supercomputing capacity. The Top500 (as of June 2022) includes five Brazilian supercomputers. Three out of these five belong to the same oil and gas company and are dedicated to seismic applications. A major performance issue with stencils is their bad data locality, which leads to memory inefficiency. This BEPE project aims to study code transformations and optimization for improving the efficiency of stencil codes that simulate the propagation of the acoustic wave across NUMA architecture. (AU)