Intel多核与集成众核上CFD程序的OpenMP性能分析

发布时间：2018-04-19 01:25

本文选题：多核 + 集成众核　；参考：《计算机科学与探索》2015年10期

【摘要】：多核与众核已成为当前主流的高性能计算体系结构,Open MP编程是开发其并行计算能力的主要手段之一。针对一个实际高阶精度结构网格CFD(computational fluids dynamics)应用程序,采用基于硬件计数器的性能测试和模型分析的方法,系统地研究了其在Intel Xeon E5 Sandy Bridge多核处理器和Intel Knights Corner集成众核协处理器上的Open MP性能。重点分析了Open MP库开销、线程负载均衡性、主存访问带宽对性能的影响,发现因Open MP并行引入的冗余计算对并行效率影响很小,但串行计算部分和负载不均衡性对并行效率影响大,主存访问带宽对浮点性能的影响大。还比较了该程序两种体系结构上的性能差异,讨论了性能进一步优化的方向。
[Abstract]:OpenMP programming is one of the main methods to develop its parallel computing capability.For an actual high-order precision structured grid CFD(computational fluids dynamic program, the method of performance testing and model analysis based on hardware counter is adopted.The performance of Open MP on Intel Xeon E5 Sandy Bridge multi-core processor and Intel Knights Corner integrated multi-core coprocessor is systematically studied.The effects of Open MP library overhead, thread load balance and main memory access bandwidth on performance are analyzed. It is found that redundant computing introduced by Open MP parallel has little effect on parallel efficiency.However, the parallel efficiency is greatly affected by the unbalance of the serial computing part and the load, and the main memory access bandwidth has a great influence on the floating-point performance.The performance differences between the two architectures are compared and the direction of further performance optimization is discussed.
【作者单位】：国防科技大学并行与分布处理重点实验室;国防科技大学计算机学院;
【基金】：国家自然科学基金~~
【分类号】：TP332

【相似文献】