基于国产众核处理器三维地震声波正演模拟

3D seismic acoustic wave forward modeling based on domestic heterogeneous many-core processors

三维地震声波理论与计算方法是地质勘探研究的基础,通过分析不同介质中声波的传播特性,完成三维地震声波正演模拟。针对三维地震声波有限差分交错网格方程正演过程中存在数值计算大、内存消耗大等实际问题,提出了基于神威·太湖之光超级计算机系统中国产异构众核处理器(申威26010)的三维地震声波正演模拟编程模型,完成了基于处理器间的进程级并行基于计算核心间的线程级并行优化策略。研究了DMA(直接内存读取)通信方式,提出2.5D流水线任务划分、通信与计算的相互掩盖的多角度优化策略。实验结果表明,该策略有效缓解了带宽瓶颈,发挥了处理器强大的计算能力,解决了程序在申威26010异构众核处理器处理有限差分问题时,并行效率过低的问题。在大规模测试下,使用266240个计算核心,程序仍能够保持稳定的计算性能,达到5.5 GFlops的场值更新。

3D seismic acoustic wave theory and calculation method are the basis of geological exploration research.By analyzing the propagation characteristics of acoustic waves in different media,it applied the 3D seismic acoustic wave forward mode-ling in exploration work.In order to solve the problem of huge numerical calculation and large memory consumption while proceeding 3D seismic wave equation staggered grid finite difference forwarding model,this paper studied and implemented the parallel optimization on the heterogeneous many-core processors of the Sunway TaihuLight super computer.Based on the implementation of a two-level parallel programming model by using MPI+Sunway Athread,it generated the DMA communication,2.5D pipelining task division and other optimization strategies.The model with this improvement reduced the negative effects by bandwidth and greatly utilized the computing power.In large-scale conditions,it tackled the issue of low efficiency about program execution on SW26010 heterogeneous many-core processors.The experimental results reveal that the performance of parallelism of a single node is much better than that of a master core.Another example is the calculation of solving numerical stress is 80 times faster on the core group than that on the single master core.This experiment can keep a constant performance by using 266240 cores,5.5 GFlpos updates of field.