异构计算环境下的三维Kirchhoff叠前深度偏移混合域并行算法

Hybrid domain parallel algorithm for 3D Kirchhoff prestack depth migration in the heterogeneous environment

三维Kirchhoff叠前深度偏移(KPSDM)面对数据量的不断增长以及可编程图形处理器(GPU)的引入,传统的并行策略已经不再适合当前的超大规模异构集群的体系架构。本文提出了一种新的混合域KPSDM并行算法,从成像空间、输入数据两个维度对偏移任务进行拆分,消除了任务之间的依赖性。为了应对异构计算环境,将核心计算部分移植到GPU上,并实现了"动态异步"的任务调度策略,保证了负载均衡;对于KPSDM执行过程中重复访问地震数据与旅行时场带来的巨大I/O开销,利用作业节点的本地存储构建分布式缓存系统,解决KPSDM可扩展性受限于共享存储能力的问题。在256节点的集群上处理实际地震数据,获得了接近线性的加速比效果。

The 3 D Kirchhoff prestack depth migration(KPSDM)is the most important depth-domain imaging method in the seismic data processing.Currently seismic data size of a single survey exceeds100 TB,and will increase to more than 1 PB in the near future.Considering the continuous increasing survey sizes and the introduction of programmable graphic process unit(GPU),the conventional parallel strategy is no longer appropriate for the largescale heterogeneous processing clusters.In this paper,we propose a practical hybrid domain parallel KPSDM algorithm based on two-level decomposition including imaging space and seismic data.The algorithm eliminates the dependency among tasks.In a heterogeneous environment,we implement the computing part on GPU and design adynamic and asynchronous" task allocation policy to achieve load balancing on heterogeneous computing system.Because KPSDM,as part of its execution,usually requires repeated access to huge seismic data and a large amount of travel time tables,the scalability is always limited by the shared storage max-imum throughput.To solve the scalability problem,we build a distributed cache system using the local storage for a KPSDM job spans.It can provide a very high aggregate data bandwidth to supply seismic data and travel time table to a running task timely.The KPSDM implementation can obtain close to linear speedup when it processes real seismic data on a 256-node cluster.