Research On Thermal-aware Management Techniques Based On System-Level

With the rapid development of semiconductor technology and increasing integration of the integrated circuit chip, power/energy is a critical resource in all types of computing systems from servers to embedded system. Energy costs dominate data center expenses and carbon footprints, system’s battery life limits the embedded devices’ functionality. However, increasing temperature of all components affect system performance, reliability seriously, and exacerbate consume energy/power. How to control power/energy of computer system (or called "thermal-aware management technology") is one of the key issues concerned by academia and industry.Operating system controls resource allocation and task scheduling in computer system, which can combine hardware technologies of dynamic voltage frequency scaling (DVFS) and dynamic power management (DPM) with runtime information of system to conduct thermal-aware management.This dissertation aims to balance system performance, fairness, power and temperature and so on, combine operating system and architecture theory, study key theories and technologies of characteristics of the task behavior, according group managing threads and retain components in the same status and so on. And realize thermal-aware task and resource management at operating system.This dissertation focuses on how to improve thermal efficiency of the CPU, memory and the whole system, and the proposed solutions are as follows:1. Summarize current research results of relative fields, analyze shortcomings of the existing power and temperature management technologies;2. Propose frequency affinity, which retain the same frequency of the CPU as long as possible, reducing frequently switching CPU frequency, improving CPU power efficiency. And based on the different CPU demands of the threads we partition them into groups, scheduling threads according groups. Coordinate frequency affinity with thread group scheduling to conduct thermal-aware management of the CPU;3. Propose memory affinity, which improves locality of accessing memory, low power temporarily unused memory rank, improving memory power efficiency of the memory. Although, modify page allocation method of the operating system, aggregating memory of one thread into one memory rank, and partitioning threads of allocating the same memory rank into the same group. Coordinate memory affinity with thread group scheduling to conduct thermal-aware management of the memory;4. Analyze the difficult of combining both CPU and memory power/temperature management. And based on thread group, combining frequency affinity and memory affinity, conduct thermal-aware management of the whole system;5. Through analyzing characteristics of the task behavior, firstly propose the concept of the "function event", portray the hot and cold characteristics of the thread and dependencies among threads, abstract task pattern according to function event. Conduct system level thermal-aware management based on function event;6. Analyze how the current thermal-aware management technologies impact the operating system in detail, and propose solutions to these problems.The contributions and innovations of our works include:1. Expand the concept of "thread group", partition threads into groups according task behavior, allocate CPU and memory resource and so on, and use frequency affinity and memory affinity, design thermal-aware task scheduling and resource allocation algorithm, realize system level thermal-aware control under the premise of performance and fairness;2. Propose a new page allocation method of aggregating one thread’s memory into the same memory rank, combine thread group scheduling, prolong access the same memory rank and others idle time, reduce memory power as much as possible;3. Firstly propose system level "function event" analysis, use function event to portray the characteristics of the task behavior, direct thermal-aware management. Function event is atomic, can better portray the nature of the task behavior, concluding hot and cold characteristics and inter-thread dependencies. It’s meaningful to complete methods of task behavior analysis and promote the theory of system level power/temperature optimization;...