Extremely Fast and Energy Efficient One-way Wave Equation Migration on GPU-based heterogeneous architecture

Abstract

One-way Wave Equation Migration (OWEM) is a classic seismic imaging method offering a good trade-off between quality and compute cost in most geological cases. In recent years, GPU-based heterogeneous architecture has gained popularity for seismic imaging. In this paper, we present a generic design for asynchronous processing and data management. By applying this design, we present an efficient GPU implementation of OWEM combining OpenACC and CUDA. Our approach improves upon classic designs by exploring asynchronous compute and data transfer between CPU and GPU using high-speed NVLink, completely masking the cost of MPI communications and I/O. Using 3, 01S GPUs, our fine-tuned OWEM can process 11, 172 seismic shots in less than 75 minutes. By tuning CPU and GPU clock frequencies, we achieve around 30% energy saving with only 4% loss of performance on PANGEA III supercomputer. We believe our design combined with the energy-aware tuning will be beneficial to many GPU applications.