基于拟态计算的大数据高效能平台设计方法  

Design method of big data high-efficiency platform based on mimic computing

作　　者：李斌 周清雷[2] 斯雪明[1] 聂凯 Li Bin;Zhou Qinglei;Si Xueming;Nie Kai(State Key Laboratory of Mathematical Engineering & Advanced Computing, Information Engineering University, Zhengzhou 450001, China;School of Information Engineering, Zhengzhou University, Zhengzhou 450001, China)

机构地区：[1]信息工程大学数学工程与先进计算国家重点实验室,郑州450001 [2]郑州大学信息工程学院,郑州450001

出　　处：《计算机应用研究》2019年第7期2059-2064,共6页Application Research of Computers

基　　金：国家重点研发计划资助项目(2016YFB0800100,2016YFB0800101);国家自然科学基金资助项目(61250007);国家“863”计划资助项目(2009AA012201)

摘　　要：针对当前大数据应用主要以通用处理器为计算核心,且系统结构单一、能效比低,无法充分满足大数据的计算需求,基于拟态计算模型,提出了一种大数据高效能平台的设计方法。以算粒为基本研究对象,深入剖析大数据应用算法的特征,合理划分各计算子任务;其次,构造体系结构匹配矩阵,将子任务分配到合理的处理部件上;最后,利用动态电压/频率调节技术和数据布局算法实现非关键任务的电压控制,并优化关键任务的结构布局。实验结果表明,拟态计算能深度融合各异构计算部件,建立具有灵活、可拓展的体系结构,充分发挥系统整体执行效率,降低功耗,提高能效比。In view of the current big data applications mainly use the general processor as the computing core, and the system structure is simple, energy efficiency ratio is low, can't fully meet the big data computing needs. Based on mimic computing model, this paper put forward a design method of big data high-efficiency platform. This method took computing grain as the basic research object, deeply analyzed the features of big data application algorithms, and reasonably divided the computatio- nal subtasks. Secondly, it constructed an architecture matching matrix and assigned the subtasks to the right processing units. Finally, it used dynamic voltage/frequency scaling technology and data layout algorithm to control the voltage of non-critical tasks, and optimized the structure layout of critical tasks. The experimental results show that the mimic computing can integrate the heterogeneous computing components in depth, establish a flexible and scalable architecture,and give full play to the overall efficiency of the system, reduce the power consumption and improve the energy efficiency ratio.

关 键 词：大数据 拟态计算 算粒 匹配矩阵 能效比 

分 类 号：TP301.4[自动化与计算机技术—计算机系统结构]