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

发布时间：2018-02-25 17:01

本文关键词： 实时功耗模型性能计数器多元线性回归绿色计算　出处：《计算机学报》2014年03期 　论文类型：期刊论文

【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).
[Abstract]:The real-time acquisition of power consumption is a prerequisite for dynamic power management, and the performance counter can accurately reflect the use of each hardware component. In this paper, a method of building system real-time power model based on fine-grained performance counter is presented. The method consists of power model setting, performance counter selection, data sampling and model solving. The result test consists of five steps, the basic idea is: select the reflection processor, The fine granularity performance counter of the main components such as memory I / O is used as the set of preselected eigenvalues. The corresponding relation between the set and the total power consumption of the running system is established through experiments, and the multivariate linear regression method is used to solve the problem. The influence coefficients of each performance counter on power consumption are obtained, and the real-time power consumption model of the system is established, and the set of effective performance eigenvalues that have a practical impact on the power consumption of the system is obtained. The experiments on three different platforms show that, The maximum error of the power model established by this method is less than 4. The model is solved by using three different sets of sampling data for each platform, and the effective feature set is completely consistent, which proves the stability of the model. Based on the experimental results of three platforms, a set of performance counters containing 70 elements is given as the default eigenvalue set of this method.
【作者单位】：中国科学院计算技术研究所;中国科学院大学;江南计算技术研究所;国家并行计算机工程技术研究中心;
【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
【分类号】：TP332.12

【共引文献】