CPU与GPU协同并行的多分量地震数据各向异性叠前时间偏移

Multi-component anisotropy prestack time migration based on collaborative parallel computing with CPU and GPU

地震勘探工区规模的日益庞大,造成多分量地震数据的各向异性叠前时间偏移算法耗时巨大。目前常用CPU集群方式并行加速该类算法,而集群方法必将导致节点间通信时耗增大;同时受限于CPU结构特点,只能通过扩大集群规模提高加速比。针对上述问题,提出一种基于CPU与GPU协同并行的多分量地震数据各向异性叠前时间偏移算法,利用OpenMP和CUDA实现CPU与多个GPU的协同并行,使用内存映射方法降低I/O耗费;并根据CPU与GPU的结构特点优化地震数据及速度数据的读取和存储方式,分割成像空间以节省算法内存消耗,每次只计算与开启的GPU个数相同条数的主测线,主测线内部采用一个GPU线程对应一道地震数据的偏移计算方法,以充分利用GPU计算能力。应用约29G的实际工区多分量地震数据分别比较不同个数GPU协同CPU并行的加速比,得知使用6个GPU协同CPU对实际纵波及转换波数据进行并行偏移处理时,加速比分别达到444和449。

The anisotropy prestack time migration(PSTM)algorithm costs a lot of time in practical application due to current seismic survey sizes.CPU-cluster computing is usually used to accelerate PSTM,but it leads to the increase of communication cost between the computing nodes.According to the CPU structure,the only way to improve a cluster performance is to enlarge the scale of cluster.To solve this problem,a multi-component anisotropy PSTM algorithm based on collaborative parallel computing with CPU and GPU is proposed.This algorithm implements CPU collaborative parallel computing with a multi-GPUs environment and OpenMP and compute unified device architecture(CUDA).It decreases I/O costs with the memory map method,and optimizes reading and storing seismic data according to the structural features of CPU and GPU.The algorithm segments image volume to decrease memory occupation,computes the number of inlines which is equal to the number of available GPUs.One GPU thread will image one seismic data trace amongst inline.The propose approach was tested on 29 GB seismic data with different numbers of GPUs.Between P-P wave and P-S wave seismic data,speedups are 444 and 449 when parallel computing with CPU collaborative parallel computing to 6 GPUs.