基于任务阶段迁移的异构多核能效优化问题研究

发布时间：2018-01-06 16:18

本文关键词：基于任务阶段迁移的异构多核能效优化问题研究　出处：《中国科学技术大学》2017年硕士论文　论文类型：学位论文

【摘要】：异构多核处理器由多个在性能和功耗上具有巨大差异的核组成,这些差异为本文优化处理器的能效提供了可能。根据任务在执行过程中表现出的不同特征将其迁移到合适的核上执行,可以提高处理器的能效。本文中的能效指处理器运行任务时的性能功耗比。然而,要充分发掘异构多核处理器的潜力还需克服一些困难。首先,最适合执行某个任务的核会随着任务执行阶段的改变而发生变化;其次,传统估算任务在不同类型核上性能功耗比的模型准确度较低;最后,需合适的算法调度所有正在运行的任务到核的分配。针对以上三点问题,本文进行了基于任务阶段迁移的异构多核能效优化问题研究,主要研究工作包括:1).提出一种基于指令类型向量的任务执行阶段划分方法。针对常用异构处理器中硬件性能计数器数目较少的限制,本文对指令类型向量划分任务执行阶段的硬件体系结构进行了改进,改进后可以节约硬件资源和减少计算量。根据指令类型向量划分任务执行阶段的基本思想设计了划分阶段算法,实验结果表明该算法能有效地划分任务的执行阶段。2).建立性能功耗比估算模型。本文建立了用于估算任务不同执行阶段在不同类型核上性能功耗比的模型。首先分析可能影响处理器性能功耗比的事件,然后采用相关性分析的方法计算出这些事件与处理器性能功耗比的相关性,导出影响力较大的几个事件,最后以这些事件作为变量建立估算模型。实验结果表明相对于传统估算模型准确度提高5.2%。3).构建基于任务迁移的能效优化模型用于指导核间任务迁移。优化模型用于抽象出本文的能效优化问题。基于Kuhn-Munkres(KM)算法求解优化模型,根据任务不同执行阶段在不同类型核上性能功耗比,确定能效最高的迁移路径。实验结果表明整个方案可以提高处理器的能效。
[Abstract]:Heterogeneous multi-core processor is composed of a plurality of great differences in performance and power consumption of the core, provides the possibility for these differences is the optimal processor efficiency. The migration to the appropriate core according to the different characteristics of the task execution in the implementation process, can improve the processor's efficiency. The energy efficiency refers to the performance of CPU power than the task. However, to fully exploit the potential of heterogeneous multi-core processors need to overcome some difficulties. First of all, the most suitable for performing a task with the nuclear task execution stage of change; secondly, the traditional estimation task in different types of nuclear power than the accurate performance model low degree; finally, need a proper algorithm for scheduling all running tasks to nuclear distribution. To solve the above three problems, this paper carried out the task of stage migration of heterogeneous multi core optimization based on energy efficiency On the issue, the main research work includes: 1). A vector instruction type task execution stage division method based on hardware. According to the number of commonly used in heterogeneous processor performance counters with fewer limitations, this paper improves the instruction type vector partitioning task execution hardware architecture stage, improved can save hardware resource and reduce the the amount of calculation. According to the basic idea of the type of command vector partitioning task execution stage design stage of division algorithm, the experimental results show that this algorithm can effectively perform the task stage.2). Establish of power consumption than the estimation model. This paper established to estimate different tasks in the phase of the execution of different types on the performance of nuclear power than the model. First analyze the possible impact event processor performance and power consumption ratio, then by using the method of correlation analysis to calculate these events and processor performance Correlation power ratio, several large events are finally to influence these events as variables to establish estimation model. The experimental results show that compared with the traditional estimation model to improve the accuracy of 5.2%.3). The construction optimization is used to guide the inter core task migration efficiency model based on task migration. For abstract energy efficiency optimization the optimization model based on Kuhn-Munkres (. KM) algorithm to solve the optimization model, according to the different tasks in the phase of the execution of different types of nuclear power performance ratio, determine the migration path of the highest efficiency. The experimental results show that the scheme can improve the efficiency of the processor.

【学位授予单位】：中国科学技术大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP332

【相似文献】