一种基于细粒度性能计数器建立系统实时功耗模型的方法  

Build Real-Time Power Consumption Model of a Computer System Based on Fine-Grained Performance Counters

作　　者：程华[1,2,3] 陈左宁 孙凝晖[1,2] 王明扬[3] 李超[3] 

机构地区：[1]中国科学院计算技术研究所,北京100190 [2]中国科学院大学,北京100039 [3]江南计算技术研究所,江苏无锡214083 [4]国家并行计算机工程技术研究中心,北京100190

出　　处：《计算机学报》2014年第3期687-699,共13页Chinese Journal of Computers

基　　金：国家"八六三"高技术研究发展计划重大项目"神威蓝光千万亿次高效能计算机系统研制"(2010AA01A141)资助

摘　　要：对计算机系统运行时功耗值的实时获取,是进行功耗动态管理的先决条件.性能计数器能够准确地反映各硬件部件的使用情况,进而反映其功耗情况.文中提出了一种基于细粒度性能计数器建立系统实时功耗模型的方法(SysPower),该方法由功耗模型设定、性能计数器选取、数据采样、模型求解、结果检验五个步骤组成,基本思路是:选取反映处理器、内存、I/O等主要部件运行状态的细粒度性能计数器作为预选特征值集合(Fpreslct),通过实验建立该集合与运行时系统总功耗的对应关系,采用多元线性回归法进行模型求解,得到各性能计数器对功耗的影响系数,从而建立系统的实时功耗模型,并得到对该系统运行时功耗产生实际影响的有效性能特征值集合(Feffect).在三个不同平台上的实验表明,该方法建立的功耗模型最大误差小于4%,针对每个平台使用三组不同的采样数据组合进行模型求解,得到的有效特征集合完全一致,证明了模型的稳定性.最后,综合三个平台的实验情况,给出了一个包含70个元素的性能计数器集合,作为该方法的缺省特征值集合(Fcommon_effect).The value of real-time power consumption is a prerequisite for dynamic power management. Total system power consumption is the sum of individual hardware unit power consumption, the performance counters can accurately reflect the use of the corresponding hardware unit, in turn, reflect its power consumption. In this paper, we propose a method to establish real-time power consumption model, which is based on a series of fine-grained performance counters, we call it SysPower model. The method has 5 steps, which are power model setting, performance counters selection, data sampling, model solution, result examination. The basic idea is as follows： select a collection of basic performance counters as Fpr^L~, which can reflect the utilization of processor, memory, I/O, etc. By experiments, record the relationship of these performance counters and the total power consumption, and then, obtain the influence coefficient of each per- formance counter through multiple linear regression method, so as to establish a real-time power consumption model, and get collection of characteristic value （Fetlock） of real impact on the system runtime power consumption. Experimental results on three different Linux platforms show the error rate of the model is less than 4%. Using three sets of sampled data for each platform, and the effective characteristic value sets exactly the same, which prove the stability of the model. From experimental results, this article also gives the common effective performance counter collection （Fcommon_offect）, which contains 70 elements, as the default characteristic values of SysPower model.

关 键 词：实时功耗模型 性能计数器 多元线性回归 绿色计算中图法 

分 类 号：TP316[自动化与计算机技术—计算机软件与理论]