Industry-scale finite-difference elastic wave modeling on graphics processing units using the out-of-core technique

Abstract

The difference in computational power between the few- and multicore architectures represented by central processing units (CPUs) and graphics processing units (GPUs) is significant today, and this difference is likely to increase in the years ahead. GPUs are, therefore, ever more popular for applications in computational physics, such as wave modeling. Finite-difference methods are popular for wave modeling and are well suited for the GPU architecture, but developing an efficient and capable GPU implementation is hindered by the limited size of the GPU memory. I revealed how the out-of-core technique can be used to circumvent the memory limit on the GPU, increasing the available memory to that of the CPU (the main memory) instead, with no significant computational overhead. This approach has several advantages over a parallel scheme in terms of applicability, flexibility, and hardware requirements. Choices in the numerical scheme — the numerical differentiators in particular — also greatly affect computational efficiency. These factors are considered explicitly for GPU implementations of wave modeling because GPUs are special purpose with a visible architecture.

Elastodynamic full waveform inversion on GPUs with time-space tiling and wavefield reconstruction

“…To this end, we tile the four-dimensional iteration space in the temporal direction and the slowest spatial direction [53]. This amounts to updating the field(s) on a subset of the x 3 axis rather than along the full spatial extent, at any given timestep.…”

“…Nguyen et al [27] combined it with classical blocking techniques for improved cache friendliness, a concept which Yount et al [63] utilized in order to exploit the HBM present on Xeon Phi coprocessors. Specifically, seismic finite-difference modelling has seen the time-space tiled out-of-core implementation presented in [53].…”