【作者】 张冬松；

【导师】 金士尧；

【作者基本信息】 国防科学技术大学 ， 计算机科学与技术， 2012， 博士

【摘要】 现实世界的实际需求促进了实时系统的产生和发展，随着电子技术和计算机技术的飞速发展，实时系统的应用范围也在不断扩大。这些领域的发展迫切需要提供性能更强大、操作更灵活、效能比更高的微处理器系统，以满足实时应用越来越复杂化、智能化的需求。随着VLSI技术的迅猛发展与应用需求的不断提高，芯片制造商正在推出多核芯片和片上多处理器系统，多核和多处理器平台已经变得更加普遍，受到了嵌入式实时领域越来越多的关注。但是，处理器的高性能会带来高能耗。而能耗是很多嵌入式实时系统首要考虑的因素，特别是无线移动和便携式计算设备。为了满足嵌入式实时应用的进一步发展，节能实时调度技术研究已经成为当前学术和工业界共同关注的前沿领域和热门课题之一，并且随着“绿色计算”需求的提出，变得越来越重要。本文针对多核/多处理器实时系统中的能耗问题，由多处理器系统到多核处理器系统，由偶发任务模型到一般任务模型，由理想DVFS处理器模型到实际DVFS处理器多维限制模型，从以下四个方面展开研究：1.本文提出一种多处理器系统中偶发任务节能实时调度算法。从多处理器偶发任务在线节能实时调度问题出发，提出了基于偶发任务最优实时调度LRE-TL的多处理器在线节能调度算法，通过LRE-TL与独立DVFS和DPM技术相结合，根据偶发任务动态释放所导致负载不断变化的情况，在每个TL面的初始时刻和偶发任务释放时刻对所有活跃任务实现在线动态电压和频率调节。算法不仅可以保证偶发任务集的最优可调度性，而且实验结果表明其始终优于已有算法，尤其在在高负载情况下能耗节余达到30%。2.本文提出多核系统中偶发任务在线节能实时调度算法。从偶发任务动态释放问题出发，提出了一种基于TL流调度的多核在线节能实时调度算法TL-DVFS，利用LRE-TL算法实现任务调度，在每个TL面的初始时刻、偶发任务的释放时刻进行动态电压频率调节，能较好适应因偶发任务的动态释放而产生的动态负载变化。经过系统的数学分析和仿真实验，结果表明TL-DVFS算法与现有方法相比，不仅保证了偶发任务集的最优可调度性，而且实现了更多的能耗节余，尤其在高负载情况下节能效果更明显。通过考虑任务实际执行时间往往小于最坏情况执行时间的情况，进一步提出了一种基于动态松弛时间回收的节能实时调度算法DSREM，在每个TL面的提前完成时刻实现动态松弛时间回收，降低后续偶发任务的执行频率，实现更多的节能。实验结果表明，DSREM算法在任务集总负载超过某一个值之后，其节能效果始终优于现有方法，尤其随着总负载的增加，节能效果会更好。3.本文提出一种基于Global EDF的多核系统一般任务在线节能实时调度算法。针对一般任务模型，提出了一种不基于先验知识，在同构多核系统中基于GlobalEDF在线节能硬实时调度算法GEDF-OLEASA，通过引入速度调节因子，利用松弛时间，结合动态功耗管理和动态电压/频率调节技术，降低多核系统中任务的执行速度，达到实时约束与能耗节余之间的合理折中。实验结果表明，该算法适用于不同类型的片上动态电压/频率调节技术，GEDF-OLEASA算法始终比GlobalEDF算法更节能，最多可节能15~20％，最少可节能5~10%。4.本文提出一种开销敏感的最优节能实时调度算法。为解决原有最优算法在实际开销不可忽视条件下不再具有最优性问题，提出开销敏感的基于帧任务多处理器最优节能实时调度算法。该算法根据关键速度来判断系统负载情况，确定具有最低能耗值的活跃处理器个数，然后根据状态切换开销来确定最优调度序列。该算法允许实时任务在任意处理器之间迁移，计算复杂度小，易于在实时操作系统中实现。系统的数学分析证明了该算法的最优性。本文通过对多处理器系统偶发任务模型、多核系统偶发任务模型、一般任务模型及实际处理器开销模型的节能实时调度技术研究，对多核/多处理器实时系统的节能设计实现进行了有益的探索。本文的理论分析、仿真实验和评价结果表明，上述节能实时调度技术是有效的，能够有效解决多核/多处理器实时系统的能耗问题。更多还原

【Abstract】 The actual needs of the real world promote the emergence and development ofreal-time systems. Meanwhile, with the rapid development of electronic and computertechnology, the application scope of real-time systems is also expanding. To meet theincreasingly complex and intellectual requirement for real-time systems, it needs toprovide more powerful, more flexible and higher cost-efficiency microprocessorsystems.With the rapid development of VLSI technology and the great increase ofapplication requirements, the multi-core chip and multi-processors on chip architecturedomain modern commercial processors. They have attracted more and more attention ofembedded real-time systems. However, the energy consumption of multi-core systemsbecomes increasingly prominent. Energy consumption is one of the most importantfactors considered by many embedded real-time systems, especially wireless andportable devices. Energy-efficient real-time scheduling research become one of the frontdomains and popular topics and attracts more and more attention of both academic andindustrial world. Furthermore, it becomes increasingly important with the proposal of"green computation".To solve the problem of energy consumption in real-time multi-core andmultiprocessor systems, this paper introduces energy-efficient technologies into thedesign of real-time scheduling algorithms, putting forward some proposals forenergy-efficient real-time scheduling. We start the research in terms the processormodel from multiprocessor systems to multi-core systems, the task model from sporadictasks to general tasks, the DVFS processor model from ideal to practicalmultidimensional limitation.1. We put forward an energy-efficient real-time scheduling algorithm for sporadictasks in multiprocessor platforms. The proposed method is based on LRE-TL which isan optimal real-time scheduling algorithm for sporadic tasks. With the schedulingalgorithm, LRE-TL, and the energy-efficient technology, independent DVFS and DPM,the voltage and frequency of active tasks is dynamically scaled at the initial time of eachTL plane and the release time of a sporadic task in each TL plane. Our proposalguarantees the optimal feasibility of sporadic tasks. Furthermore, the experimentalresults show that compared with existing algorithms, more energy savings can beachieved in all cases, especially up to30%in the case of high workloads.2. We propose to overcome the energy consumption problem of multi-corearchitecture for sporadic tasks, including two key technologies: TL-DVFS (Time Localremaining execution plane based Dynamic Voltage Frequency Scaling) and DSREM(Dynamic Slack Reclamation based Energy-efficient Multi-core real-time scheduling). TL-DVFS is an online real-time energy-efficient scheduling algorithm based on TLplane and adopts LRE-TL to perform real-time task scheduling. The voltage andfrequency are dynamically scaled at the initial time of each TL plane and the releasetime of a sporadic task in each TL plane, which is adaptive to the dynamic workload ofsporadic tasks and obtain more energy savings. Systematic mathematical analysis andextensive simulation results show that TL-DVFS can not only guarantee the optimalfeasibility of sporadic tasks, but also achieve more energy savings in all cases,especially in the case of high workloads. In addition, DSREM is proposed afterquantifying the characteristics that the actual execution time of tasks is usually less thantheir Worst-Case Execution Time (WCET) producing lots of dynamic slack time. Themain idea of the algorithm is to reclaim dynamic slack time, during which DVFStechniques can be used to reduce the execution frequency of future tasks to reduceenergy consumption. Experimental results show that DSREM can always save moreenergy than existing algorithms when the total workload of the system exceeds athreshold, and at the same time guarantees optimal schedulability. The performance willbe improved when the total workload is increased.3. This paper develops a Global EDF-based OnLine Energy-Aware SchedulingAlgorithm (GEDF-OLEASA) for hard real-time tasks in multi-core system.GEDF-OLEASA is based on the general task model and with no priori to tasks’properties. It can reduce the execution speed of task in multi-core system, and reach areasonable compromise between real-time constraints and energy savings, as itintroduces a speed scale factor for utilizing the slack time, and combines dynamicpower management with dynamic voltage/frequency scaling techniques. The resultsshow that the algorithm can be well applied to different kinds of dynamicvoltage/frequency scaling on chip, and compared with Global EDF algorithm, it gainmore energy savings in all cases, which can improve energy savings about15%to20%at most and about5%to10%at least.4. We propose an optimal energy-efficient real-time scheduling algorithm which isoverhead-aware. Recent optimal energy-efficient real-time scheduling ignores the timeand energy overhead of processor state switching. Thus it is no more optimal in actualplatforms. For multiprocessors with independent dynamic voltage frequency anddynamic power management, we propose an optimal energy-efficient real-timescheduling algorithm for frame-based tasks. The proposed optimal algorithm determinesthe system workload cases and the number of active processor cores in tems of criticalspeed. Then we can obtain the optimal scheduling according to the switching overhead.The algorithm allows tasks to arbitrarily migrate across processors during theirexecution at the cost of small computational complexity. Furthermore, it is easy to beimplemented. Systematic mathematical analysis shows that the algorithm is optimal.In this paper, we carried on the research from energy-efficient real-time scheduling techniques for sporadic task model in multiprocessor system, sporadic task model inmulti-core system, general task model, and practical processor overhead model. Themathematical analysis, simulation experiments and evaluation results of this paper showthat these energy-efficient real-time scheduling techniques are effective, can be used tosolve energy consumption problem of multi-core and multiprocessor real-time systems.更多还原