多核系统下吞吐量与延迟敏感的并行流水调度能耗最小化问题研究

发布时间：2018-02-17 04:57

本文关键词： 嵌入式系统流应用 DVS 节能并行流水　出处：《东北大学》2013年硕士论文　论文类型：学位论文

【摘要】：嵌入式系统功能越来越强大,性能也大幅度提高,系统能耗越来越高,系统散热问题越来越突出。能耗问题已经成为制约嵌入式系统发展的瓶颈。因此,嵌入式系统的节能调度已经成为嵌入式领域一个非常重要的研究课题。在多核架构中,对流应用采取并行流水调度已越发有效并广泛流行于现有的各类系统中,如媒体和图像处理系统等。流应用通常计算密集,对能量要求很高,他们运行在能源有限的系统中时会引发很多问题,因此节能研究将显得尤为重要。本文主要研究支持片内DVS (Dynamic Voltage Scaling)技术的多核处理器平台中的流应用的并行流水调度节能问题。流应用被建模为一个加权有向无环图结构的任务图。一个流应用通过静态的并行流水调度策略并以基于流应用的有向无环图的结构被分配到多核系统中。现有的DVS技术相关研究大多基于连续的频率调节模型,然而这并不适用于实际系统。因此,本文研究的能耗模型为离散频率调节模型。本文的目标是在保证吞吐量和延迟等服务质量要求下,确定最优化的频率分配使流应用的能耗最小化。本文针对支持受限缩放技术的处理器模型,提出了2种不同的启发式频率设定算法：PPS算法和PPC算法。PPS算法使所有任务初始频率设定为最高值,然后逐渐寻找到任务降低频率时使单位时间内降低的能耗最多的任务,降低其频率即抻长其执行时间,直到所有任务均不能再抻长为止。PPC算法使所有任务初始频率设定为最低值,然后逐渐增加各任务频率即缩短任务执行时间,至所有任务满足吞吐量约束为止,然后逐渐寻找到任务增加频率时使单位时间内增加的能耗最少的任务,缩短其执行时间,直到所有任务的执行时间满足延迟的约束为止。然后将其扩展到支持任意缩放技术的处理器模型中,并将并行流水模型合理分段,基于分段并行流水模型提出了2种不同的启发式频率设定算法：PPSS算法和PPSC算法。通过模拟实验表明,本文提出的算法节能效果显著,PPSS算法和PPSC算法效果比PPS算法和PPC算法更好。
[Abstract]:The function of embedded system is more and more powerful, the performance of embedded system is greatly improved, the system energy consumption is higher and higher, the problem of system heat dissipation is more and more prominent. The energy consumption problem has become the bottleneck restricting the development of embedded system. Energy saving scheduling of embedded system has become a very important research topic in embedded field. In multi-core architecture, parallel income scheduling has become more and more effective and popular in all kinds of existing systems. Such as media and image processing systems. Stream applications are usually computationally intensive and require a lot of energy. They can cause a lot of problems when they run in systems with limited energy. Therefore, the research on energy saving will be particularly important. This paper mainly studies the problem of parallel income scheduling energy saving in multi-core processor platform supporting DVS dynamic Voltage scaling technology. The flow application is modeled as a weighted directed acyclic. The task diagram of graph structure. A stream application is assigned to multi-core system by static parallel income scheduling strategy and directed acyclic graph structure based on flow application. Most of the existing research on DVS technology is based on continuous frequency regulation model. However, this is not suitable for practical systems. Therefore, the energy consumption model studied in this paper is a discrete frequency regulation model. The purpose of this paper is to ensure the quality of service, such as throughput and latency. Determining optimal frequency allocation minimizes energy consumption for stream applications. This paper proposes two different heuristic frequency setting algorithms: PPS algorithm: PPS algorithm and PPC algorithm. PPS algorithm sets the initial frequency of all tasks to the maximum value, and then gradually finds out the task that reduces the energy consumption per unit time when the task reduces the frequency. Reduce its frequency, that is, stretch its execution time, until all tasks can no longer be stretched. PPC algorithm sets the initial frequency of all tasks to the lowest value, and then gradually increases the frequency of each task, that is, shortens the task execution time. Until all tasks meet throughput constraints, and then gradually find the task that increases the frequency of the task with the least increase in energy consumption per unit of time, shortening its execution time, Until the execution time of all tasks meets the constraints of delay. Then it is extended to a processor model that supports arbitrary scaling techniques, and the parallel income model is reasonably segmented. Based on the piecewise parallel income model, two different heuristic frequency setting algorithms: PPSC and PPSC are proposed. The simulation results show that the proposed algorithm is more efficient than PPS and PPC.
【学位授予单位】：东北大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP368.1

【参考文献】