单处理器及多处理器系统节能技术的研究

发布时间：2018-05-14 21:33

本文选题：单处理器 + 多资源　；参考：《东北大学》2012年博士论文

【摘要】：随着数字化进程的日益加剧,系统的节能问题显得越发的重要。节能技术优劣不但影响电池供电嵌入式系统的使用或工作时间,即生命周期(lifetime),而且很大程度上决定了大规模系统(例如数据中心)的电费开销。为了延长嵌入式系统的生命周期,以及降低大规模系统的运行成本,在软件和硬件层次上,大量工作对节能技术进行了研究。在硬件方面,许多处理器提供了动态电压缩放(DVS)的功能,即处理器可以工作在不同的电压／频率上。同时,几乎所有的外围设备也支持动态电源管理(DPM),即设备具有多种不同的工作状态或模式。在软件方面,通过调节系统负载(任务)在处理器上的工作状态、任务划分和调度以及使用外设的方式来达到节能的目的。随着片外设备数量的增加和多核系统的普及,多资源以及多核系统的节能问题受到了越来越多的关注。本文从软件层面上,主要对单处理器多资源和多核实时系统的节能问题进行深入的研究。首先,在单处理器系统中,根据设备和处理器的功耗在系统总体功耗中所占的不同比例,给出不同的解决方法。其次,根据多核系统不同的划分以及任务的不同特性,提出相应的解决方法。具体来说,本文工作主要包含以下几个方面： (1)在处理器功耗占系统功耗主要部分(外设功耗可忽略或是恒为常量)的单处理器系统中,考虑处理器模式切换的时间和能量开销,研究能耗敏感的实时任务调度及其可调度性测试条件。首先,提出新的可调度性测试条件,大大降低了其悲观性；其次,通过任务合并消除处理器的空闲模式,大大减少模式切换次数,从而降低功耗；最后,放松对处理器在每个协周期内对休眠时间的限制,使算法适用于更多类型的处理器。 (2)针对外设功耗处于系统整体功耗的决定性部分的嵌入式系统,研究了典型的无线传感器网络(WSN)的节能问题。无线传感器网络己经在监控系统等应用中被广泛采用,因为传感节点往往是由电量有限的电池供电,所以,如何恰当的控制每个传感节点的能量消耗从而最大化WSN的生命周期是至关重要的。本文研究输油管道监控系统中传感器节点的线性布置问题,目的是最大化该WSN的生命周期。对于简单的等距布置方式,首先,说明基于被普遍接受的理想功耗模型的一个结论(增加传感器节点可以增加WSN的生命周期)并不适用于真实的功耗模型；然后,研究等功耗放置方式,并将该问题建模为混合整数线性规划(MILP)；最后,提出两个高效的启发式算法,从相反的方向搜索各传感节点工作的功耗等级。与等距放置策略相比,提出的两个启发式算法大大的降低了系统能耗,有效地平衡了各传感节点的能耗,显著的增加了WSN的生命周期。其中一个算法的结果与MILP的最优解几乎相同。 (3)在处理器与外设功耗相当的系统中,研究单处理器多外设实时系统的节能调度问题。具体工作主要包含两个部分。首先,对于简单的基于帧的周期任务,针对连续和离散处理器频率模型,本文分别提出高效的算法,通过计算使系统运行能耗最小的处理器最优频率和设备最优空闲时间,来实现全系统节能。其次,本文研究具有固定数量实时任务和固定数量外设的系统,考虑不可忽略的设备转换时间和能耗开销,找出能耗最优调度,包括任务的执行顺序,任务的运行频率以及设备状态转换的时间点。对于不同的系统配置,分别采用数学规划结合启发式算法的方式解决该问题,实验结果表明提出的算法大大降低了系统的能耗。 (4)对于多核实时系统,研究划分为簇(cluster)或岛(island)的多核体系结构的节能调度问题,这种体系结构下,每个岛上的所有处理器(核)具有相同的工作电压和频率。该研究综合考虑了系统的时间和频率约束,对实时任务提出能耗最小化的调度算法。首先,证明在不考虑时间约束的情况下,每个岛的能耗最小化的最优频率并不依赖于映射到该岛上的负载,而是依赖于该岛的核数及其漏电功耗。然后,针对具有时间约束的系统,在固定任务划分情况下,提出一多项式复杂度的算法最小化能耗,并证明其最优性。最后,给出多项式复杂度的整体算法来确定系统活跃岛的数量,任务划分和任务频率分配。实验表明该算法在节能方面大大优于相关的方法,并且分析了不同簇划分的节能效果。 (5)研究同质多处理器／多核系统并行实时任务的节能调度问题。对于执行在固定个数处理器上的并行任务,首先,提出几个基于层装箱(level-packing)的启发式任务调度算法,大大降低各层内的空闲时间；然后,提出一个多项式复杂度能耗最小化算法,并证明其最优性。对于执行并行任务的处理器个数可以变化的情况,提出另一个多项式复杂度的算法来确定执行各个任务的处理器个数,任务调度以及任务的频率分配。实验结果表明提出的算法可以大大的降低系统的能耗。总之,本文综合研究了单处理器及多处理器系统的节能技术。首先,研究单处理器多资源系统的节能技术,对处理器功耗占主要部分的系统,外设功耗占主要部分的系统以及二者相当的系统,分别提出了不同的节能方法；然后,对于多核实时系统,研究了划分为岛的多核系统以及并行实时任务的节能调度问题,大大的降低了系统的能耗。
[Abstract]:With the intensification of the digital process, the energy saving problem of the system becomes more and more important. The advantages and disadvantages of energy saving technology not only affect the use or working time of the battery powered embedded system, that is, the life cycle (lifetime), but also largely determine the cost of electricity in large-scale systems (such as the data center). The life cycle, as well as reducing the operating cost of large-scale systems, studies energy saving technology at the software and hardware levels. On the hardware side, many processors provide the function of dynamic voltage scaling (DVS), that is, the processor can work on different electric voltage / frequency. At the same time, almost all peripheral devices are also supported. Dynamic power management (DPM), that is, a device has a variety of different working states or modes. In software, the purpose of energy saving is achieved by adjusting the working state of the system load (task) on the processor, task division and scheduling, and using peripherals. With the increase of the number of devices and the popularization of multi core systems, many resources are available. The energy saving problem of multi-core system has attracted more and more attention.
From the software level, this paper mainly studies the energy saving problem of single processor multi resource and multi core real-time system. First, in the single processor system, different solutions are given according to the different proportion of the power consumption of the device and the processor in the system overall power consumption. Secondly, according to the different division of the multi-core system, Specific solutions to the different characteristics of tasks are proposed. Specifically, the work of this paper includes the following aspects:
(1) in a single processor system whose power consumption is the main part of the system power consumption (peripheral power consumption is negligible or constant), the time and energy cost of processor mode switching is considered, and the energy sensitive real-time task scheduling and its schedulability test conditions are studied. First, the new schedulability test conditions are proposed, which greatly reduce the test conditions. Pessimism; secondly, eliminating the idle mode of the processor by merging the task, greatly reducing the number of mode switching times and reducing the power consumption; finally, the relaxation time limit for the processor in each co cycle is relaxed, so that the algorithm is suitable for more types of processors.
(2) the energy saving of the typical wireless sensor network (WSN) is studied in the embedded system which is the decisive part of the power consumption in the whole system. The wireless sensor network has been widely used in the application of the monitoring system, because the sensor nodes are often powered by the limited battery, so how to control properly The energy consumption of each sensor node and thus maximizing the life cycle of WSN is essential. This paper studies the linear arrangement of sensor nodes in the pipeline monitoring system to maximize the life cycle of the WSN. For a simple isometric arrangement, first, one is based on a widely accepted ideal power model. The conclusion (increasing the sensor node can increase the life cycle of WSN) is not suitable for the real power consumption model. Then, we study the power placement method and model the problem into mixed integer linear programming (MILP). Finally, two efficient heuristic algorithms are proposed to search the power consumption level of each sensor node from the opposite direction. Compared with the equidistant placement strategy, the proposed two heuristic algorithms greatly reduce the energy consumption of the system, effectively balance the energy consumption of each sensor node, and significantly increase the life cycle of WSN. One of the algorithms is almost the same as the optimal solution of MILP.
(3) the energy saving scheduling problem of the single processor multi peripheral real-time system is studied in the system with the equal power consumption of the processor and peripherals. The specific work includes two parts. First, for the simple frame based periodic task, the efficient algorithm is proposed for the continuous and discrete processor frequency model, and the system runs through calculation. The optimal frequency of the processor and the optimal idle time of the equipment to achieve the whole system energy saving. Secondly, this paper studies a system with fixed number of real-time tasks and fixed number of peripherals, and considers the time and cost of energy consumption that can not be ignored, to find the optimal adjustment of energy consumption, including the sequence of task execution, and the frequency of the task. As well as the time point of the device state conversion, the problem is solved by mathematical programming and heuristic algorithm for different system configuration. The experimental results show that the proposed algorithm greatly reduces the energy consumption of the system.
(4) for multi-core real-time systems, study the energy saving scheduling problem divided into cluster (cluster) or island (Island) multi-core architecture. Under this architecture, all the processors on each island have the same working voltage and frequency. The study takes into consideration the time and frequency constraints of the system, and minimizes energy consumption for real-time tasks. First, it is proved that the optimal frequency of energy consumption minimization of each island does not depend on the load mapped to the island without considering the time constraints, but depends on the number of nuclei and the power leakage of the island. Then, for a system with time constraints, a polynomial complexity is proposed in the case of a fixed task division. The algorithm minimizes energy consumption and proves its optimality. Finally, the overall algorithm of polynomial complexity is given to determine the number of active islands, task division and task frequency distribution. The experiment shows that the algorithm is much better than the related methods in energy saving and analyses the energy saving effect of different cluster division.
(5) to study the energy saving scheduling problem of parallel real-time tasks in homogeneous multiprocessor / multi-core systems. For parallel tasks on a fixed number processor, first, several heuristic task scheduling algorithms based on layer packing (level-packing) are proposed, which greatly reduce the idle time in each layer; then, a polynomial complexity energy consumption is proposed. Minimize the algorithm and prove its optimality. For the change in the number of processors that perform parallel tasks, another polynomial complexity algorithm is proposed to determine the number of processors that perform each task, task scheduling and the frequency allocation of tasks. The experimental results show that the proposed algorithm can greatly reduce the energy consumption of the system.
In conclusion, the energy saving technology of single processor and multi processor system is studied in this paper. Firstly, the energy saving technology of the single processor multi resource system is studied, the system with the main part of the power consumption, the system with the main part of the peripheral power consumption and the two equivalent system, the different energy saving methods are put forward, and then, for the multi kernel, the multi core energy saving technology is put forward. In real time systems, the multi-core system divided into islands and the energy saving scheduling problem of parallel real-time tasks are studied, which greatly reduces the energy consumption of the system.

【学位授予单位】：东北大学
【学位级别】：博士
【学位授予年份】：2012
【分类号】：TP332

【共引文献】