基于内外子区域划分的高阶精度CFD程序异构并行算法

对计算流体力学(CFD)程序CNS提出一种Offload模式下基于内外子区域划分的异构并行算法，结合结构化网格下有限差分计算和四阶龙格库塔方法的特点，引入ghost网格点区域，设计了一种ghost区域收缩计算策略，显著降低了异构计算资源之间的数据传输开销，负载均衡时CPU端的计算与MPI通信完全和加速器端的计算重叠，提高了异构协同并行性。推导了保证计算正确性的ghost区域的参数，分析了负载均衡的条件。在“CPU(Intel Haswell Xeon E5-2670 12 cores ×2)＋加速器(Xeon Phi 7120A ×2)”的服务器上测得该算法较直接将任务子块整体迁至加速器端计算的异构算法性能平均提升5.9倍，较MPI/OpenMP两级并行算法使用24个纯CPU核的性能，该算法使用单加速器时加速1.27倍，使用双加速器加速1.45倍。讨论和分析了性能瓶颈与存在的问题。

A heterogeneous parallel algorithm based on inner-out subdomain dividing for high order CFD solver

An Offload-Mode heterogeneous parallel algorithm based on inner-out subdomain dividing was proposed for Computational Fluid Dynamics (CFD) program CNS. Combined with the characteristics of finite difference computing and fourth order Runge-Kutta method in structure mesh, ghost region was introduced, based on which a Ghost-Region-Shrinking computing scheme was designed, significantly reducing the overhead of data movement between heterogeneous computing resources, making the computing and MPI communication on CPU absolutely overlap with the accelerator computing under load balance condition, bringing better heterogeneous synergetic parallelism. Parameter of the ghost region for the computing validity was given and load balance tuning was demonstrated. On a server with CPU (Intel Haswell Xeon E5-2670 12 cores ×2) + MIC (Xeon Phi 7120A ×2), an averaged performance improvement of 5.9x was gained over the algorithm of using accelerator with task blocks integrally. Compared with MPI/OpenMP two-level parallel algorithm running on 24 Intel Haswell CPU cores, an accelerating of 1.27x was achieved with one MIC and 1.45x with two MICs. The bottleneck and disadvantage were discussed.