一种改进的整机系统实时功率建模方法  

Improved Methodology for Full-system Power Modeling

作　　者：杨良怀[1] 戚加欣 徐卫[1] 范玉雷[1] YANG Liang-huai;QI Jia-xin;XU Wei;FAN Yu-lei(School of Computer Science and Technology,Zhejiang University of Technology,Hangzhou 310023,China)

机构地区：[1]浙江工业大学计算机科学与技术学院

出　　处：《小型微型计算机系统》2019年第7期1503-1509,共7页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61702456,61502420,61070042)资助

摘　　要：本文通过考察处理器、内存和磁盘三个部件的活动信息,包括硬件性能事件和部件利用率,构建整机系统实时功率模型.对于处理器和内存的功耗,通过特征选择找到合适的处理器内部的性能事件,结合CPU功耗状态C-States,建立低误差的功耗模型;对于没有提供性能事件的磁盘,通过利用率对部件功耗进行估算.采用回归方法分别建立并综合两个模型即可得到整机的实时功率模型.实验使用了真实系统Eureqa建模和PosgreSQL系统运行TPC-H负载对两种功率模型进行验证,结果表明:基于硬件性能事件与C-States的实时处理器与内存功率模型的误差在3%以下;结合性能事件、C-States与利用率的整机功率模型其相对误差在4%~10%之间,优于仅使用利用率的整机功率模型,最大时提升可达7%.This paper aims to constructs full-system power models which are independent of the running workloads.Our models exploit the activity information of the main components(CPU,disk,and memory),including performance events and device utilization.When modeling power consumption for processor and memory,by combining performance events with the processor’s C-States,we obtained a power model with small error.To model disk’s power consumption,we estimated its power by its utilization since there are no diskrelated performance events for use till now.These two models are constructed with the regression method and are combined to get the power model of the whole machine.We compared our models with those previously proposed by using workloads from Eureqa and TPC-H benchmark queries in PosgreSQL.Experimental results showed that the average relative errors of our power model for processor and memory is less than 3%,and similarly that average relative errors of our power models for full-system are between 4%and10%,which sometimes reach as high as 7%better than models that only exploits device utilization.

关 键 词：功率建模 软功率计 性能计数器 性能事件 

分 类 号：TP303[自动化与计算机技术—计算机系统结构]