三维全波形反演高效异构并行计算

摘要

全波形反演(FWI)每次迭代都需要进行若干次地震波正演,计算量非常大,尤其在三维情况下,提高并行计算的效率和稳健性至关重要.引入随机边界来反传、重建震源波场,可充分发挥GPU的计算能力,从而实现反演梯度的高效计算,相比监测点(checkpoint)和有效边界技术,大幅减少了数据存储和数据交换的开销,具有计算效率高和存储量小的优点;开发了作业池并行作业管理机制,与常规消息传递接口(message passing interface,MPI)并行机制相比,可动态增减节点,具有近似线性的加速比,更适应大规模异构并行.采用三维SEG/EAGE推覆体模型进行了速度反演测试,结果证明该技术高效且可靠.%Each iteration of full waveform inversion (FWI) requires several seismic forward modeling and takes large computational effort especially in 3D cases.It is key to improve the efficiency and robustness of parallel computing for FWI.To maximize the ability of GPU,we introduce random boundary to back propagate source wavefield for reconstruction to efficiently compute the gradient inversion.Compared to checkpoints and effective boundary technology,it can significantly reduce the cost of data storage and exchange.To adapt to massive heterogeneous parallel clusters,we develop a parallel job management mechanism named job pool.Compared to the conventional message passing interface (MPI) mechanism,it can add in or remove out computing nodes on the fly and the speedup ratio is approximately linear.Numerical example of 3D Overthrust model for velocity inversion showed high efficiency and reliability of the method.