基于MIC集群平台的GMRES算法并行加速

发布时间：2018-05-07 19:10

本文选题：广义极小残量法 + MIC　；参考：《计算机科学》2017年04期

【摘要】：广义极小残量法(GMRES)是最常用的求解非对称大规模稀疏线性方程组的方法之一,其收敛速度快且稳定性良好。Intel Xeon Phi众核协处理器(MIC)具有计算能力强、易编程、易移植等特点。采用MPI+OpenMP+offload混合编程模型将GMRES算法移植到MIC集群平台上。采用进程间集合通信异步隐藏、数据传输优化、向量化以及线程亲和性优化等多种手段,大幅提升了GMRES算法的求解效率。最后将并行算法应用到"局部径向基函数求解高维偏微分方程"问题的求解中。测试表明,CPU节点集群上开启32个进程,并行效率高达71.74%,4块MIC卡的最高加速性能可达单颗CPU的7倍。
[Abstract]:Generalized minimal residue method (GMRES) is one of the most commonly used methods for solving asymmetric large sparse linear equations. Its convergence speed is fast and its stability is good. The Intel Xeon Phi multikernel coprocessor MICM has the advantages of strong computing power, easy programming and easy transplantation. MPI OpenMP offload hybrid programming model is used to transplant GMRES algorithm to MIC cluster platform. Asynchronous hiding, data transmission optimization, vectorization and thread affinity optimization are used to improve the efficiency of GMRES algorithm. Finally, the parallel algorithm is applied to the problem of solving high dimensional partial differential equations with local radial basis function. The test results show that 32 processes are opened on the cluster of CPU nodes, and the parallel efficiency is as high as 71.74 / 4 MIC cards. The maximum acceleration performance of the four MIC cards is 7 times that of a single CPU.
【作者单位】：浪潮集团高效能服务器和存储技术国家重点实验室;太原理工大学数学学院;
【分类号】：O246

【相似文献】