基于时间—能耗权重比的任务调度算法
发布时间：2019-03-08 16:36
【摘要】：异构并行系统是高性能低功耗计算机系统的主要发展趋势之一,异构并行系统下的低功耗研究是近年来学者研究的热点问题。由于动态电压频率调整(DVFS,Dynamic Voltage and Frequency Scaling)技术以及任务调度在能耗优化方面的潜能,近年来不少基于DVFS技术的任务调度算法被提出并取得了不错的效果。然而,这些算法没有综合考虑任务的执行时间和能耗,无法做到时间和能耗的权衡优化,且这些算法在其任务模型中使用平均值表示任务的执行时间和通信时间,这种表示方法不够精确,会对任务调度的性能产生影响。异构并行系统下基于时间-能耗权重比的任务调度算法(Weight-ratio-based Task Scheduling,WTS)与DVFS技术相结合,它同时考虑时间和能耗这两个性能指标,能够根据时间-能耗的权重比为每个任务选择合适的处理器及电压级别,做到时间和能耗的权衡优化,获取相较于现有算法更优的系统加权性能。在任务初次分配阶段,WTS算法为每个任务选择使其加权性能提升值最大的处理器及电压级别;在任务再次优化分配阶段,该算法随机选择一个任务,并将该任务重新分配给相对于原有分配有系统加权性能提升的处理器及电压级别,使系统的加权性能进一步提升。同时,WTS算法考虑到任务执行时间以及任务间通信时间的不确定性,在任务模型中使用近似权重代替平均值,以获取更优的调度性能。为证明WTS算法的有效性,仿真实验将其与两个现有算法在时间、能耗、系统加权性能以及相对性能提升等方面进行了对比,实验结果显示,WTS算法能做到时间和能耗的权衡优化,使系统的加权性能更优,同时在时间和能耗单个性能指标方面也具有优势。
[Abstract]:Heterogeneous parallel systems are one of the main trends in the development of high performance and low power computer systems. The research on low power consumption in heterogeneous parallel systems is a hot issue in recent years. Due to the potential of dynamic voltage-frequency adjustment (DVFS,Dynamic Voltage and Frequency Scaling) and task scheduling in energy optimization, many DVFS-based task scheduling algorithms have been proposed and achieved good results in recent years. However, these algorithms do not consider the task execution time and energy consumption, and can not achieve the balance of time and energy consumption optimization, and these algorithms use the average in their task model to represent the execution time and communication time of the task. This representation is not accurate enough to affect the performance of task scheduling. In heterogeneous parallel systems, the task scheduling algorithm (Weight-ratio-based Task Scheduling,WTS) based on time-to-energy weight ratio is combined with DVFS technology, which takes into account both time and energy consumption. It can select the appropriate processor and voltage level for each task according to the weight ratio of time-to-energy consumption, and optimize the trade-off of time and energy consumption, and obtain the better system weight performance than the existing algorithms. In the initial assignment phase, the WTS algorithm selects the processor and voltage level whose weighted performance increases the maximum for each task. In the task reoptimization phase, the algorithm randomly selects a task, and reassigns the task to the processor and voltage level, which improves the weighted performance of the system compared with the original assignment, which further improves the weighted performance of the system. At the same time, considering the uncertainty of task execution time and inter-task communication time, WTS algorithm uses approximate weight instead of average value in task model to obtain better scheduling performance. In order to prove the effectiveness of the WTS algorithm, the simulation experiment compares it with the two existing algorithms in terms of time, energy consumption, system weighting performance and relative performance improvement. The experimental results show that: The WTS algorithm can optimize the balance of time and energy consumption, and make the weighted performance of the system better. At the same time, it also has advantages in the single performance index of time and energy consumption.
【学位授予单位】：华中科技大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP301.6
本文编号：2436992
[Abstract]:Heterogeneous parallel systems are one of the main trends in the development of high performance and low power computer systems. The research on low power consumption in heterogeneous parallel systems is a hot issue in recent years. Due to the potential of dynamic voltage-frequency adjustment (DVFS,Dynamic Voltage and Frequency Scaling) and task scheduling in energy optimization, many DVFS-based task scheduling algorithms have been proposed and achieved good results in recent years. However, these algorithms do not consider the task execution time and energy consumption, and can not achieve the balance of time and energy consumption optimization, and these algorithms use the average in their task model to represent the execution time and communication time of the task. This representation is not accurate enough to affect the performance of task scheduling. In heterogeneous parallel systems, the task scheduling algorithm (Weight-ratio-based Task Scheduling,WTS) based on time-to-energy weight ratio is combined with DVFS technology, which takes into account both time and energy consumption. It can select the appropriate processor and voltage level for each task according to the weight ratio of time-to-energy consumption, and optimize the trade-off of time and energy consumption, and obtain the better system weight performance than the existing algorithms. In the initial assignment phase, the WTS algorithm selects the processor and voltage level whose weighted performance increases the maximum for each task. In the task reoptimization phase, the algorithm randomly selects a task, and reassigns the task to the processor and voltage level, which improves the weighted performance of the system compared with the original assignment, which further improves the weighted performance of the system. At the same time, considering the uncertainty of task execution time and inter-task communication time, WTS algorithm uses approximate weight instead of average value in task model to obtain better scheduling performance. In order to prove the effectiveness of the WTS algorithm, the simulation experiment compares it with the two existing algorithms in terms of time, energy consumption, system weighting performance and relative performance improvement. The experimental results show that: The WTS algorithm can optimize the balance of time and energy consumption, and make the weighted performance of the system better. At the same time, it also has advantages in the single performance index of time and energy consumption.
【学位授予单位】：华中科技大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP301.6
【参考文献】
相关期刊论文 前3条
1 林一松;杨学军;唐滔;王桂彬;徐新海;;一种基于关键路径分析的CPU-GPU异构系统综合能耗优化方法[J];计算机学报;2012年01期
2 PADUA Divid;;Communication contention in APN list scheduling algorithm[J];Science in China(Series F:Information Sciences);2009年01期
3 刘必慰;陈书明;汪东;;先进微处理器体系结构及其发展趋势[J];计算机应用研究;2007年03期
相关博士学位论文 前2条
1 王桂彬;大规模异构并行系统软件低功耗优化关键技术研究[D];国防科学技术大学;2011年
2 陈娟;低功耗软件优化技术研究[D];国防科学技术大学;2007年
相关硕士学位论文 前1条
1 张泰忠;基于遗传算法的异构多核关键任务低功耗调度研究[D];湖南大学;2011年
，本文编号：2436992
本文链接：https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/2436992.html
