CPU-GPU协同的Andorid平台游戏功耗管理研究
发布时间：2018-09-12 06:06
【摘要】：当前移动智能终端设备数量增长迅猛,性能提升显著,在满足用户需求的同时,凸显了其固有的能耗与电池容量之间的矛盾。动态电压频率调节(DVFS)是能够有效解决这种矛盾的一种机制。但是传统策略中,处理器负载评估依赖底层硬件反馈、CPU和GPU功耗管理模块完全独立,无法充分利用不同类型应用的独有特性进一步提高移动智能终端设备的能效。本文关注Android智能手机处理器功耗管理和能效提升。以"上层应用特性驱动的底层处理器功耗管理"思想为指引,选取Android移动平台游戏应用为研究对象,探究游戏运行时行为特性和CPU-GPU负载特性,指导CPU-GPU协同功耗管理方案的设计与实现。具体来说,本文的主要工作包括:1.构建了针对Android平台游戏应用的功耗管理框架。框架避免对PC平台的过度借鉴,利用操作系统的应用管理和资源分配功能,纵向联通应用层和底层处理器功耗管理模块,使DVFS算法可以直接感知上层游戏计算需求变化;横向联通CPU和GPU相互独立的功耗管理模块,使二者可以相互感知对方进行协同管理。框架为基于游戏特性的CPU/GPU负载评估和CPU-GPU协同DVFS算法奠定基础。2.提出用户交互感知的CPU负载评估方法和基于游戏帧复杂度的GPU负载评估方法。方法利用Android游戏运行时的用户交互特性,以用户交互消息数评估CPU负载中的用户交互部分;利用CPU-GPU之间的数据交互特性,在CPU处以游戏中顶点、纹理和命令数据提前评估GPU负载,解决了传统基于硬件利用率反馈方法延迟高和准确度低的问题。3.设计了针对Android游戏的CPU-GPU协同DVFS算法。算法以保证用户体验为导向,对游戏用户体验、CPU-GPU负载构建了变化感知的协同描述,为电压频率调控指明时机和方向;协同调频逻辑基于游戏运行时CPU-GPU之间功能上的生产者——消费者协作关系设计,做到"协同"而非"同时独立"的CPU-GPU功耗管理。本文在ODROID-XU3实验平台上实现了上述工作。实验评价中,本文以平均帧处理时间(性能)和平均帧能耗(能耗)的比值作为游戏运行时系统的能效量化值。实验结果表明:相比传统方法,本文方案在保证Android游戏用户体验前提下可以达到平均12%,最高34%的能效提升。
[Abstract]:At present, the number of mobile intelligent terminal devices is growing rapidly and the performance is improving significantly. While meeting the needs of users, it highlights the contradiction between its inherent energy consumption and battery capacity. Dynamic voltage frequency regulation (DVFS) is an effective mechanism to solve this contradiction. But in the traditional strategy, the processor load evaluation depends on the underlying hardware feedback CPU and GPU power management module completely independent, can not make full use of the unique characteristics of different types of applications to further improve the energy efficiency of mobile intelligent terminal devices. This paper focuses on Android smart phone processor power management and energy efficiency. Guided by the idea of "the power management of the bottom processor driven by the upper application characteristics", this paper selects the Android mobile platform game application as the research object, and probes into the game runtime behavior characteristics and the CPU-GPU load characteristics. To guide the design and implementation of CPU-GPU collaborative power management scheme. Specifically, the main work of this paper includes: 1. 1. A power management framework for Android platform game applications is constructed. The framework avoids the excessive reference to the PC platform and uses the application management and resource allocation functions of the operating system to connect the application layer and the underlying processor power management module vertically so that the DVFS algorithm can directly perceive the changes of the upper level game computing requirements. The power management module of CPU and GPU is independent of each other so that they can perceive each other and manage each other collaboratively. The framework lays the foundation for CPU/GPU load evaluation based on game features and CPU-GPU cooperative DVFS algorithm. 2. 2. The CPU load evaluation method based on user interaction awareness and the GPU load evaluation method based on game frame complexity are proposed. Methods the user interaction part of the CPU load was evaluated by the number of user interactive messages in the Android game run time, and the vertex of the CPU was determined by the data interaction property between the CPU-GPU. Texture and command data are used to evaluate GPU load in advance, which solves the problem of high delay and low accuracy based on traditional feedback method of hardware utilization. A CPU-GPU cooperative DVFS algorithm for Android games is designed. Under the guidance of ensuring user experience, the algorithm constructs a cooperative description of change perception for CPU-GPU load of game user experience, which indicates the timing and direction of voltage frequency regulation. The cooperative frequency modulation logic is based on the producer-consumer cooperative relationship between the CPU-GPU when the game is running, and achieves "collaborative" rather than "simultaneous independent" CPU-GPU power management. In this paper, the above work is realized on the ODROID-XU3 experimental platform. In the experimental evaluation, the ratio of average frame processing time (performance) and average frame energy consumption (energy consumption) is used as the energy efficiency quantization value of the game runtime system. The experimental results show that: compared with the traditional method, this scheme can achieve an average of 12 and a maximum of 34% energy efficiency improvement under the premise of guaranteeing the Android game user experience.
【学位授予单位】：中国科学技术大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP311.52;TP316
本文编号：2238130
[Abstract]:At present, the number of mobile intelligent terminal devices is growing rapidly and the performance is improving significantly. While meeting the needs of users, it highlights the contradiction between its inherent energy consumption and battery capacity. Dynamic voltage frequency regulation (DVFS) is an effective mechanism to solve this contradiction. But in the traditional strategy, the processor load evaluation depends on the underlying hardware feedback CPU and GPU power management module completely independent, can not make full use of the unique characteristics of different types of applications to further improve the energy efficiency of mobile intelligent terminal devices. This paper focuses on Android smart phone processor power management and energy efficiency. Guided by the idea of "the power management of the bottom processor driven by the upper application characteristics", this paper selects the Android mobile platform game application as the research object, and probes into the game runtime behavior characteristics and the CPU-GPU load characteristics. To guide the design and implementation of CPU-GPU collaborative power management scheme. Specifically, the main work of this paper includes: 1. 1. A power management framework for Android platform game applications is constructed. The framework avoids the excessive reference to the PC platform and uses the application management and resource allocation functions of the operating system to connect the application layer and the underlying processor power management module vertically so that the DVFS algorithm can directly perceive the changes of the upper level game computing requirements. The power management module of CPU and GPU is independent of each other so that they can perceive each other and manage each other collaboratively. The framework lays the foundation for CPU/GPU load evaluation based on game features and CPU-GPU cooperative DVFS algorithm. 2. 2. The CPU load evaluation method based on user interaction awareness and the GPU load evaluation method based on game frame complexity are proposed. Methods the user interaction part of the CPU load was evaluated by the number of user interactive messages in the Android game run time, and the vertex of the CPU was determined by the data interaction property between the CPU-GPU. Texture and command data are used to evaluate GPU load in advance, which solves the problem of high delay and low accuracy based on traditional feedback method of hardware utilization. A CPU-GPU cooperative DVFS algorithm for Android games is designed. Under the guidance of ensuring user experience, the algorithm constructs a cooperative description of change perception for CPU-GPU load of game user experience, which indicates the timing and direction of voltage frequency regulation. The cooperative frequency modulation logic is based on the producer-consumer cooperative relationship between the CPU-GPU when the game is running, and achieves "collaborative" rather than "simultaneous independent" CPU-GPU power management. In this paper, the above work is realized on the ODROID-XU3 experimental platform. In the experimental evaluation, the ratio of average frame processing time (performance) and average frame energy consumption (energy consumption) is used as the energy efficiency quantization value of the game runtime system. The experimental results show that: compared with the traditional method, this scheme can achieve an average of 12 and a maximum of 34% energy efficiency improvement under the premise of guaranteeing the Android game user experience.
【学位授予单位】：中国科学技术大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP311.52;TP316
【参考文献】
相关期刊论文 前1条
1 帖军;陈幼均;;进程同步中的生产者—消费者模型分析[J];武汉科技学院学报;2007年04期
，本文编号：2238130
本文链接：https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/2238130.html
