基于神经网络的独立程序在单机上运行功耗的预测  

Power Analysis for Executable Program on Single Computer Based on Artificial Neural Network

作　　者：谭一鸣[1] 曾国荪[2] 

机构地区：[1]同济大学计算机科学与技术系,上海200092 [2]国家高性能计算机工程技术中心同济分中心,上海200092

出　　处：《计算机科学》2012年第5期282-286,294,共6页Computer Science

基　　金：863项目(2009AA012201);973计划课题(2007CB316502);国家自然基金项目(90718015);NSFC-微软亚洲研究院联合资助项目(60970155);教育部博士点基金项目(20090072110035);上海市优秀学科带头人计划项目(10XD1404400);高效能服务器和存储技术国家重点实验室开放基金项目(2009HSSA06)资助

摘　　要：程序运行能耗分析是目前绿色计算、高效能计算的研究热点。由于应用程序自身的复杂性、硬件平台的异构性、环境因素对功耗影响的不确定性,导致很难直接给出程序运行功耗的预测公式,因此提出一种基于神经网络的程序运行功耗预测方法。通过构造一个BP神经网络,以影响程序运行功耗的软件、硬件和环境因素为神经网络的输入,以程序运行功耗和执行时间为神经网络的输出,并选取多个应用领域的典型实用程序来提取相应特征参数,且将其作为训练样本来训练神经网络,从而得到程序功耗预测模型。重复实验表明,提出的功耗预测方法具有合理性和可行性。Power management of application program is a hot research in the area of green computing and high productivity computing.Because of the complexity of application programs,heterogeneity of processors and uncertainty of the running environment,it is difficult to propose an accurate method to predict energy consumption for application program directly.So we presented a power analysis paradigm for application program based on artificial neural network.First,we built a power analysis model based on back propagation neural network（BPNN）.The three factors of software,hardware and environment were taken as the inputs of BPNN,and energy consumption and finish time as the outputs of BPNN.Next,we chose lots of classic application programs from different fields as training samples.After learning and training,an expected BPNN was obtained which can be used to predict energy consumption for other new programs.Repeated experiments show that this power analysis paradigm is rational and feasible.

关 键 词：绿色计算 功耗分析 神经网络 可执行程序 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]