高性能计算机功耗管理系统设计与实现

发布时间：2019-03-18 11:17

【摘要】：高性能计算机在能源、生物、气象、科研、地质勘探等计算密集型应用中得到长足发展，是一个国家科技综合实力的重要象征，在现代社会中发挥着重要的作用，然而它的发展现在却面临着功耗墙的重大挑战。随着半导体工艺的持续发展和芯片集成度的显著提高，微处理器的性能在得到大幅度提升的同时也造成了功耗密度的急剧增大，而且伴随着计算机结点规模的扩大，高性能计算机的功耗急剧增高，目前千万亿次的计算机功耗基本以兆瓦（MW）计，其中K Computer功耗更是达到了惊人的9.898MW。高功耗同时带来高发热，消耗了更多的冷却成本，造成巨大的电力压力，大大限制了高性能计算机的发展和应用。本文以高性能计算机功耗管理为研究内容，基于对系统负载、能耗分布以及硬件特征等总体把握的基础上，设计功耗管理系统，该系统集监控、策略制定、功耗调整、设备控制于一体，结合资源管理、作业管理、低功耗编译、动态电源管理、动态电压调整以及外围设备控制等方法技术，力求在保证系统性能的同时降低系统的整体功耗。在功耗管理系统的研究中，自主开发了模拟器Simschedule对ParallelWorkloads Archive标准负载进行分析，，总结出高性能计算机作业负载的特点。建立结点能耗模型，结合作业负载特点，提出了三种基于关闭空闲结点的低功耗调整算法，三种算法分别为关闭间隔时间调整算法、反比关系时间调整算法、历史记录调整算法，通过模拟验证的方式证明了三种算法的有效性。在满足用户切换频率限定约束的情况下，三种算法都大大降低了系统的功耗。本文以天河-1A超级计算机为实验平台对外围设备的功耗管理进行了研究。经过分析，基于CPU负载、CPU温度以及风机转速之间存在的相互关系，提出了基于感知的风机智能调控策略，在达到精确制冷的同时，能够降低风机22.9%的功耗；对电源单元提出了一种按需分配的供电策略，在保证电源单元时时处于高效率工作的前提下，达到节省功耗17.5%的效果。采用功耗管理系统较好地实现了对高性能计算机系统级功耗管理和对外围设备的调控，大大降低了整个计算机系统的整体功耗，节约了经济成本，同时具有良好的社会效应。
[Abstract]:High-performance computer has made great progress in computing intensive applications such as energy, biology, meteorology, scientific research, geological exploration and so on. It is an important symbol of the comprehensive strength of science and technology in a country and plays an important role in modern society. However, its development is now facing major challenges from the power-consuming wall. With the continuous development of semiconductor technology and the remarkable improvement of chip integration, the performance of microprocessor not only improves greatly, but also results in a sharp increase of power density, and accompanied by the expansion of computer node size. The power consumption of high-performance computers has increased dramatically. At present, the power consumption of thousands of terabytes of computers is basically megawatt (MW), and the K-Computer power consumption has reached an astonishing 9.898MW. High power consumption at the same time leads to high heat consumption, consuming more cooling costs, resulting in huge power pressure, which greatly limits the development and application of high-performance computers. Based on the overall grasp of system load, energy consumption distribution and hardware characteristics, this paper designs a power management system based on high-performance computer power management. The system includes monitoring, policy formulation, power consumption adjustment, and so on. Combined with resource management, job management, low-power compilation, dynamic power management, dynamic voltage regulation and peripheral equipment control, equipment control can ensure the performance of the system and reduce the overall power consumption of the system. In the research of power management system, the simulator Simschedule is developed independently to analyze the ParallelWorkloads Archive standard load, and the characteristics of high performance computer job load are summarized. According to the characteristics of job load, three low-power adjustment algorithms based on closing idle nodes are proposed. The three algorithms are closed interval time adjustment algorithm and inverse ratio relation time adjustment algorithm, respectively. The validity of the three algorithms is verified by simulation. The three algorithms greatly reduce the power consumption of the system when the limit of user switching frequency is satisfied. In this paper, the power management of peripheral devices is studied with Tianhe-1A supercomputer as the experimental platform. Based on the analysis of the relationship among CPU load, CPU temperature and fan speed, an intelligent fan control strategy based on perception is proposed, which can reduce the power consumption of the fan by 22.9% while achieving accurate refrigeration. This paper presents a power supply strategy based on demand distribution for the power supply unit, which can save 17.5% of power consumption on the premise that the power supply unit is in high efficiency all the time, and that the power consumption of the power supply unit can be reduced by 17.5%. The power management system is adopted to realize the power management of high performance computer system and the control of peripheral devices, which greatly reduces the overall power consumption of the whole computer system, saves the economic cost, and has good social effects at the same time.
【学位授予单位】：国防科学技术大学
【学位级别】：硕士
【学位授予年份】：2012
【分类号】：TP38

【参考文献】