• 中国计算机学会会刊
  • 中国科技核心期刊
  • 中文核心期刊

J4 ›› 2016, Vol. 38 ›› Issue (05): 848-856.

• 论文 • 上一篇    下一篇

异构多核下兼顾应用公平性和能耗的调度方法研究

杨亚琪,栾钟治,杨海龙,杨姝,钱德沛   

  1. (北京航空航天大学计算机学院中德软件技术联合研究所,北京 100191)
  • 收稿日期:2015-12-26 修回日期:2016-02-15 出版日期:2016-05-25 发布日期:2016-05-25
  • 基金资助:

    国家自然科学基金国际合作与交流项目(61361126011);国家863计划(2012AA010904)

Application fairness and energy aware
scheduling on heterogeneous multi-core processors   

YANG Yaqi,LUAN Zhongzhi,YANG Hailong,YANG Shu,QIAN Depei   

  1. (SinoGerman Joint Software Institute,School of Computer Science and Engineering,Beihang University,Beijing 100191,China)
  • Received:2015-12-26 Revised:2016-02-15 Online:2016-05-25 Published:2016-05-25

摘要:

异构多核处理器通常由高性能的大核和低能耗的小核组成,在其上进行合理的线程调度可以有效地提高资源利用率,节省能耗。之前论文提出的大小核上的公平性调度并没有考虑核上有不同频率/电压状态的情况,而现在支持DVFS调节的处理器越来越普遍,因此很有必要将线程间公平度的计算进行扩展和改进。提出在每个核有若干种不同的DVFS状态时异构多核处理器上线程公平度的计算方法,对已有的性能预测模型进行改进,采用自适应算法调整模型中的系数,并在此基础上提出了一种调度策略,维持各线程之间的公平度和处理器功率满足提前设定的阈值,同时选取能效最优化的配置,实现减小应用运行能耗的目的。实验结果表明,与所提出的调度策略相比,采用static、DVFSonly、swaponly三种调度方法时,在总的运行时间几乎相同的情况下,平均要多产生20%以上能耗,对于有些应用甚至达到了50%。

关键词: 异构多核, DVFS, 公平性, 能耗, 线程调度

Abstract:

Heterogeneous multicore processors mostly consist of highperformance big cores and energyefficient small cores. Proper scheduling on such kind of processors can improve resource utilization and reduce energy consumption effectively. Fairnessaware scheduling proposed previously does not take DVFS into account. However, processors supporting DVFS become more and more common. It is necessary to extend fairnessaware scheduling to processors with multiple frequency/voltage states on each core. In this paper, we propose a scheduling on heterogeneous multicore processors which can reduce the energy consumption of the workload by keeping both the fairness of all threads and the total power of the processor satisfying the preset threshold. We propose a calculation method for the fairness among cores with multiple frequency/voltage states, and improve the existing performance prediction models which can estimate the performance of a thread for all configurations using hardware counters collected on the current configuration. Then the best scheduling policy is chosen, which can not only satisfy the fairness and the power threshold, but also maximize the energy efficiency at the same time. Some experiments are done with the Odroid XU3 development board and more experiments with Sniper simulator. Experimental results show that compared with the static, DVFSonly, swaponly scheduling, the proposal can reduce energy consumption obviously while the total running time is almost not affected.

Key words: heterogeneous multicores;DVFS;fairness;energy consumption;thread scheduling