地震处理多核异构并行计算通用框架研究  

The research of universal seismic processing frame on heterogeneous multi-core parallel

作　　者：孟祥宾[1] 隋志强[1] 王修银[1] 唐祥功[1] 段疾病 

机构地区：[1]胜利油田分公司物探研究院,山东东营257022 [2]中国海洋大学信息科学与工程学院,山东青岛266100

出　　处：《油气地球物理》2014年第2期11-16,共6页Petroleum Geophysics

摘　　要：新的地震采集方式可每天产生30TB的巨大数据,迫切需要并行计算技术支撑资料处理,而并行模式的复杂化(如MPI、OpenMP和CUDA等)导致程序设计的复杂化,尤其当系统软硬件资源变化时,必须反复修改源程序。为简化复杂的地震并行软件开发,提升地震处理并行效率,本文在分析各种并行模式的基础上,建立了一整套地震处理多核异构并行计算通用框架,将各种并行模式相结合,充分发挥各自的优势,实现地震处理软件的多核异构并行模式自动匹配,提高了地震处理软件在多核异构环境下开发的可行性和并行效率。基于该框架研发的GPU炮域波动方程叠前深度偏移软件,与CPU串行算法相比,计算精度等同,但并行效率提高20倍以上,且随GPU节点增多呈线性增长趋势。New seismic acquisition can generate enormous volume of data, about 30TB per day, which resorts to parallel computation for supporting data proccssing on time. However, complex parallel computation pattern （such as MPI, OpenMP, CUDA） makes program desing more difficult, especially when repeated modification of source program is necessary due to system software and hardware resources variation. To facilitate seismic parallel software development and enhance its operation efficiency,a universal seismic processing frame on heterogeneous parallel multi-core architecture is designed based on analyses of various parallel patterns. Its key characteristic is automatic pattem matching by integrating various parallel modes to fully play their respective advantages, which also greatly improves parallel development feasibility and efficiency. A shot domain seismic wave equation pre-stack depth migration module is developed on this frame which runs 20 times faster than its serial CPU algorithm in the same precision and shows linear increment as more GPU nodes are used.

关 键 词：地震处理 多核异构 并行模式 通用框架 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]