有限元网格积分算法在MIC众核平台上的并行实现

发布时间：2018-05-31 07:21

本文选题：集成众核 + 卸载模式　；参考：《计算机科学》2015年11期

【摘要】：基于英特尔集成众核(Many Integrated Core,MIC)架构,将有限元网格积分算法在至强融核(Xeon Phi)协处理器做了移植和性能分析。该应用全面测试了有限元分析的核心计算过程在MIC上的加速效果,实现了卸载模式(offload)[1]下利用OpenMP在MIC上的线程并行化。计算性能测试结果显示集成众核平台可以有效地加速有限元网格积分算法:1)一块被充分利用的MIC设备卡(3115A)的计算能力超过两路16核Intel XeonTM E5-2670CPU;2)MIC并发的物理线程可能由于公共缓存访问存在竞争而降低程序的扩展性。测试结果还显示了在多CPU多MIC平台上进一步移植完整的MPI并行有限元模拟软件的可行性。这项工作有助于推动与有限元网格相关的科学和工程高性能计算的研究。
[Abstract]:Based on the Many Integrated Core (MIC) architecture, the finite element mesh integration algorithm has been transplanted and performance analyzed in the most strong fusion kernel (Xeon Phi) coprocessor. The application tests the acceleration effect of the core calculation process of the finite element analysis on MIC and realizes the use of OpenMP in MIC under the unloading mode (offload) [1]. The results of the computational performance test show that the integrated public kernel platform can effectively accelerate the finite element mesh integration algorithm: 1) a fully utilized MIC device card (3115A) has more computing power than two 16 core Intel XeonTM E5-2670CPU; 2) MIC concurrent physical threads may be reduced due to the competition of public cache access. The test results also show the feasibility of further transplantation of the complete MPI parallel finite element simulation software on multiple CPU multiple MIC platforms. This work helps to promote the research of scientific and engineering high performance computing related to the finite element grid.
【作者单位】：上海超级计算中心;埃克塞特大学数学系;
【基金】：国家高技术研究发展计划(863)(2012AA01A308) 国家自然科学基金(11473014) 上海市科学技术委员会科研计划项目(13DZ2294500)资助
【分类号】：O241.82;TP338.6

【参考文献】