基于BP神经网络的处理器节能技术研究
Research on Processor Energy Saving Strategy Based on BP Neural Network
作者：郭兵(四川大学 计算机学院, 四川 成都 610065)；张鹏博(四川大学 计算机学院, 四川 成都 610065)；沈艳(成都信息工程大学 控制工程学院, 四川 成都 610054)；黄义纯(四川大学 计算机学院, 四川 成都 610065)；曹亚波(四川大学 计算机学院, 四川 成都 610065)
Author：GUO Bing(College of Computer Sci., Sichuan Univ., Chengdu 610065, China)；ZHANG Pengbo(College of Computer Sci., Sichuan Univ., Chengdu 610065, China)；SHEN Yan(School of Control Eng., Chengdu Univ. of Info. Technol., Chengdu 610054, China)；HUANG Yichun(College of Computer Sci., Sichuan Univ., Chengdu 610065, China)；CAO Yabo(College of Computer Sci., Sichuan Univ., Chengdu 610065, China)
收稿日期：2017-03-09 年卷（期）页码：2018,50(1):107-112
期刊名称：工程科学与技术
Journal Name：Advanced Engineering Sciences
关键字：节能;DVFS;BP神经网络;动态电压与频率调节
Key words：energy saving;DVFS;BP neural network;dynamic voltage and frequency scaling
基金项目：国家自然科学基金重点项目资助（61332001）；国家自然科学基金资助项目（61772352；61472050）；四川省科技计划资助项目（2015GZ0103）；成都市科技惠民技术研发项目资助（2014-HM01-00326-SF）
中文摘要
研究芯片功耗中动态功耗部分，针对传统动态节能技术动态电压与频率调节（dynamic voltage and frequency scaling，DVFS）技术未能考虑预测CPU未来阶段行为的不足，提出BP-DVFS节能策略。为了提高下一阶段CPU利用率的预测准确性，更准确地对CPU进行动态调频进而降低其运行功耗。构建了一种FPU-CPU（forward predict utilization CPU）模型。模型假设下一时间段CPU利用率与CPU运行资源有关的事件特征量存在非线性函数关系，从处理器运行时环境出发提取出与CPU资源紧密相关的5个特征量进行度量，采用BP神经网络进行拟合训练。用训练后得到的神经网络预测CPU下一阶段的利用率，进行CPU处理不同类型任务程序的功耗仿真实验。并在相同实验条件下与常用的3种CPU调频策略实验结果进行对比。实验结果表明，在CPU处理不同类型任务程序时，采用BP-DVFS策略进行调频的CPU功耗都低于其他3种策略进行调频的CPU功耗。通过实验验证，本文提出的方法提高了预测CPU利用率的准确度，降低了CPU运行时功耗。同时验证了假设的合理性与有效性以及此方法实现CPU低功耗运行是有效的。
英文摘要
Focusing on dynamic power consumption of the chip,this paper proposed a new energy-saving strategy named BP-DVFS.Currently,DVFS (dynamic voltage and frequency scaling) technology failed to predict future behavior of CPU.In order to improve the performance of CPU utilization and predict more accurate CPU dynamic frequency,it is essential to reduce its operating power consumption.A CPU model named FPU-CPU (forward predict utilization) was built in strategy.A non-linear relationship between the CPU utilization in next period and the CPU running resource was assumed in the model.Five events related to CPU resources and extracted from run-time processor environment were measured as features.Afterwards,the training phase of was performed by the BP (back propagation) neural network.Various experiments were then carried out to predict different kinds of tasks including CPU utilization,processing and power consumption through BP neural network.To evaluate the performance of proposed method,the experimental results under same situation were compared with three traditional CPU frequency modulation strategies.The experimental results showed that the CPU power consumption of BP-DVFS based strategy is lower than that of other three strategies when CPU handles different kinds of program tasks.The experimental verification indicated that the proposed technique wasnot only accurate inpredicting performance of CPU utilization,but also helpful to reduce power consumption of CPU.
【关闭】
