起伏地表叠前时间偏移的多级并行优化技术

Kirchhoff积分法叠前时间偏移并行实现方案除了要考虑数据体规模外，还必须考虑大规模节点部署条件下的并行执行效率。而在目前的数据处理规模及集群配置条件下，利用多进程进行成像域并行的实现策略存在内存占用率高、不能充分发挥并行文件系统优势等问题，特别是在多节点大规模计算时，进程间的网络通讯、同步开销及I/O问题严重影响并行效率。为了提高起伏地表叠前时间偏移算法在大规模部署条件下的偏移效率，分析了该算法的技术特色和计算特点，提出了一种适用于积分法偏移的多级并行计算架构，采用偏移距域的多进程数据域并行、地震数据I/O与偏移计算的异步并行以及单进程内的多线程成像域并行技术对原始算法进行了架构升级及算法优化。利用34个节点的SMP集群系统和Panasas并行集群存储系统对优化前、后的算法进行测试，某工区实际叠前地震数据的测试结果表明，该并行架构能够有效降低网络通讯和数据I/O对计算性能的影响，使算法的并行效率和大规模集群环境下部署的可扩展能力同时得到提升。

The parallel implementation for Kirchhoff integral prestack time migration should adapt to large-scale data and support large-scale PC clusters.The original algorithm of Kirchhoff migration is implemented with a simple MPI parallel framework by imaging domain decomposition.The disadvantages of the parallel framework include high occupancy for memory resources,low efficiency of parallel storage,lots of synchronization operations among the processes especially with large-scale compute nodes,and therefore it does not satisfy mass seismic data and PC clusters now.In order to achieve a large-scale deployment of prestack time migration from rugged topography,we proposed an optimized multi-level parallel framework which is suitable to Kirchhoff integral migration.With techniques of multi-processes parallelism in common offset domain and the asynchronous parallelism of I/O and computation,as well as the multi-threads parallelism in imaging domain,the original algorithm is optimized and the test results demonstrate that the new parallel framework can improve the efficiency and scalability.