系统级动态电源管理框架与在线策略的研究与实现
发布时间：2018-07-20 17:36
【摘要】：随着嵌入式系统的高速发展,高性能和低功耗的矛盾日益突出,功耗已成为嵌入式系统设计的主要问题。高功耗消耗增加冷却成本,降低系统的可靠性和电池的使用期限。减少电能的消耗不仅能延长电池的寿命,降低用户更换电池的周期,而且能带来提高系统性能与降低系统开销的好处,甚至能起到保护环境的作用。低功耗设计就是在保证性能约束的条件下最大限度地降低功耗。 动态电源管理(Dynamic Power Management, DPM)已被证明是一种有效减少系统功耗的系统级低功耗技术。它在一定的请求服务和性能的约束下,对系统进行动态配置,关闭闲置系统组件或将这些组件转入低能耗状态,从而实现对能耗的有效利用。 随着对DPM的深入研究,一个标准化、支持不同电源管理策略的策略框架日益变得重要。DPM框架从整个系统的角度来处理系统的电源管理问题,采用结构化的规则和机制来整合系统不同组件的DPM技术或相关策略。 现有的DPM框架缺乏对策略的调度方法及对策略性能的有效评估,鉴于此,本文在原有DPM框架的基础上,加入策略选择与策略评估模块,对原有框架进行扩展,得到一个更为完善的DPM框架。 DPM策略是系统级动态电源管理技术研究的重点,它决定了空闲组件何时关闭或转入低功耗状态,动态电源管理对能耗的有效利用,很大程度上取决与所采用的策略的性能。传统DPM策略主要有Timeout策略、预测策略、随机模型策略三种,这些策略都有各自的缺点且策略的有效性严重依赖于一个精确的负载模型,然而对于一个复杂的系统来说,负载通常是不可预测的,因为它取决与应用程序的性质、输入数据于用户上下文。 本文在分析传统DPM策略不足的基础上,提出系统级动态电源管理的在线学习策略。策略通过奖励／惩罚改变系统状态的行为,调整行为,从而实现动态电源管理策略的在线自更新,解决在部分可观察环境(负载模型无法预测)下的动态电源管理问题。 最后,在扩展的DPM框架上,对策略选择、在线策略与传统DPM策略进行对比实验,验证了本文扩展的动态电源管理框架的有效性及策略选择和在线策略的省电性能。
[Abstract]:With the rapid development of embedded systems, the contradiction between high performance and low power consumption has become increasingly prominent, power consumption has become the main problem of embedded system design. High power consumption increases cooling costs and reduces system reliability and battery life. Reducing the consumption of electric energy can not only prolong the battery life, reduce the cycle of battery replacement, but also improve the system performance and reduce the cost of the system, and even play a role in protecting the environment. Low power design is to minimize power consumption while ensuring performance constraints. Dynamic Power Management (DPM) has been proved to be a system level low power technology which can effectively reduce system power consumption. It dynamically configures the system under the constraints of request service and performance, closes the idle system components or turns them into a low energy consumption state, so as to realize the effective utilization of energy consumption. With the in-depth study of DPM, a standardized policy framework supporting different power management strategies has become increasingly important. DPM framework deals with the power management problems of the system from the perspective of the whole system. Adopt structured rules and mechanisms to integrate DPM technology or related strategies of different components of the system. The existing DPM framework lacks the scheduling method of the policy and the effective evaluation of the policy performance. In view of this, this paper adds the policy selection and policy evaluation module on the basis of the original DPM framework, and extends the original framework. Get a more perfect DPM framework. DPM strategy is the focus of system-level dynamic power management technology. It determines when idle components are turned off or into low-power state, and dynamic power management makes effective use of energy consumption. Much depends on the performance of the strategy used. There are three kinds of traditional DPM strategies: timeout strategy, prediction strategy and stochastic model strategy. These strategies have their own shortcomings and their effectiveness depends heavily on an accurate load model. However, for a complex system, the effectiveness of these strategies depends heavily on an accurate load model. The load is usually unpredictable because it depends on the nature of the application and inputs data into the user context. Based on the analysis of the deficiency of traditional DPM strategy, this paper proposes an online learning strategy for dynamic power supply management at system level. The strategy changes the behavior of the system state and adjusts the behavior through reward / punishment to realize the online self-updating of the dynamic power management strategy and to solve the dynamic power management problem under the partially observable environment (the load model can not be predicted). Finally, in the extended DPM framework, the comparison of policy selection, online strategy and traditional DPM strategy is carried out to verify the effectiveness of the extended dynamic power supply management framework and the power saving performance of the strategy selection and online strategy.
【学位授予单位】：华南理工大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP303.3
本文编号：2134251
[Abstract]:With the rapid development of embedded systems, the contradiction between high performance and low power consumption has become increasingly prominent, power consumption has become the main problem of embedded system design. High power consumption increases cooling costs and reduces system reliability and battery life. Reducing the consumption of electric energy can not only prolong the battery life, reduce the cycle of battery replacement, but also improve the system performance and reduce the cost of the system, and even play a role in protecting the environment. Low power design is to minimize power consumption while ensuring performance constraints. Dynamic Power Management (DPM) has been proved to be a system level low power technology which can effectively reduce system power consumption. It dynamically configures the system under the constraints of request service and performance, closes the idle system components or turns them into a low energy consumption state, so as to realize the effective utilization of energy consumption. With the in-depth study of DPM, a standardized policy framework supporting different power management strategies has become increasingly important. DPM framework deals with the power management problems of the system from the perspective of the whole system. Adopt structured rules and mechanisms to integrate DPM technology or related strategies of different components of the system. The existing DPM framework lacks the scheduling method of the policy and the effective evaluation of the policy performance. In view of this, this paper adds the policy selection and policy evaluation module on the basis of the original DPM framework, and extends the original framework. Get a more perfect DPM framework. DPM strategy is the focus of system-level dynamic power management technology. It determines when idle components are turned off or into low-power state, and dynamic power management makes effective use of energy consumption. Much depends on the performance of the strategy used. There are three kinds of traditional DPM strategies: timeout strategy, prediction strategy and stochastic model strategy. These strategies have their own shortcomings and their effectiveness depends heavily on an accurate load model. However, for a complex system, the effectiveness of these strategies depends heavily on an accurate load model. The load is usually unpredictable because it depends on the nature of the application and inputs data into the user context. Based on the analysis of the deficiency of traditional DPM strategy, this paper proposes an online learning strategy for dynamic power supply management at system level. The strategy changes the behavior of the system state and adjusts the behavior through reward / punishment to realize the online self-updating of the dynamic power management strategy and to solve the dynamic power management problem under the partially observable environment (the load model can not be predicted). Finally, in the extended DPM framework, the comparison of policy selection, online strategy and traditional DPM strategy is carried out to verify the effectiveness of the extended dynamic power supply management framework and the power saving performance of the strategy selection and online strategy.
【学位授予单位】：华南理工大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP303.3
【参考文献】
相关期刊论文 前5条
1 夏军 ,郑建宏;嵌入式系统的动态电源管理架构[J];单片机与嵌入式系统应用;2005年01期
2 王暹辉;高级配置和电源接口ACPI标准介绍[J];电子技术应用;1998年12期
3 雷霆;李曦;周学海;;低能耗软件设计中的性能无损电压调度技术研究[J];计算机研究与发展;2006年06期
4 许海燕,卢伟,杨晨晖,吴芸;低能耗嵌入式系统的设计[J];计算机应用研究;2002年06期
5 陈洪明;;嵌入式系统的低功耗技术研究[J];实验科学与技术;2008年01期
相关博士学位论文 前1条
1 卜爱国;嵌入式系统动态低功耗设计策略的研究[D];东南大学;2006年
相关硕士学位论文 前2条
1 吴泽祥;系统运行时无线网卡动态电源管理的研究与实现[D];华南理工大学;2011年
2 梁晶;嵌入式系统低功耗设计研究与实现[D];华中科技大学;2004年
，本文编号：2134251
本文链接：https://www.wllwen.com/kejilunwen/jisuanjikexuelunwen/2134251.html
