Multi-GPU based two-level acceleration of full waveform inversion

Full waveform inversion (FWI) of seismic data is very computationally expensive. In this paper, we have developed a two-level parallel scheme to speed up FWI with multiple graphics processing units (multiple GPUs). The first level parallelism is the coarse-grained parallelism among multiple GPUs, which is used to reduce the number of shots; the second level parallelism is the fine-grained parallelism within each individual GPU grid, which is used to speed up the wavefield propagation procedures. The PML boundary condition is used, and the efficient boundary storage strategy is used to avoid the tremendous storage requirement needed on the disk and the data transfer between the disk and memory. We tested the scheme on the INSPUR TSI0000 system with 10 Tesla C2050 GPUs by reconstructing the Marmousi velocity model using FWI in the time domain and compared the computation time with that on CPUs and on single GPU, the result showed that the two-level based FWI is about 500 times faster than the CPU-based implementation, and the speedup of the two-level scheme is a product of those of the two levels individually. With this scheme, the turnaround time of FWI has been reduced significantly.