A Curvilinear-Cartesian Hybrid Grid 3D Finite-Difference Elastic Wave Simulation Method Based on Multi-GPU Platform

Abstract The finite-difference method with a capability of addressing undulating surfaces stands out as a popular and effective numerical simulation tool in seismological research. The computational efficiency of the wavefield simulation, a data-level parallel computing task, is severely compromised by the message passing interface (MPI)-based point-to-point boundary data interaction implemented by the traditional central processing unit (CPU) parallelism. To enhance computing power and data interaction efficiency, we have implemented a curved grid finite-difference method using multi-GPU (graphics processing unit) parallel processing with CPU-to-GPUs one-to-many mapping. Adopting a Master–Slave parallel architecture, we devised a two-stage data communication strategy and utilized GPU-direct data transmission to replace the traditional intranode data transmission that requires host intervention, thereby reducing the frequency and volume of internode data transmission. Besides, we have developed an optional hybrid grid implementation that will apply to commonly used box-shaped simulation domain with one undulating surface. It uses Cartesian grids in deep regions, reducing numerical computational complexity and memory consumption. We validated our method by examining relative differences between seismograms calculated by our method and those calculated by seismic waves fourth order (SW4) and SPECFEM3D, after normalizing by their peak amplitudes. Numerical tests have demonstrated the program’s accuracy (<3% amplitude discrepancy comparing with semi-analytical solutions), capability to address undulating surfaces (<1.2% amplitude discrepancy compared to spectral-element method solutions and 2% compared to SW4 solutions using the same grid spacing), and lower runtime (by 15% when reducing the volume of curved grids by 80%), indicating its potential as an effective tool for local seismic-wave propagation simulation.