计算机工程

• 体系结构与软件技术 • 上一篇    下一篇

功耗受限情况下多核处理器能效优化方案

邱晓杰,安虹,陈俊仕,迟孟贤,金旭   

  1. (中国科学技术大学 计算机科学与技术学院,合肥 230027)
  • 收稿日期:2016-04-15 出版日期:2017-04-15 发布日期:2017-04-14
  • 作者简介:邱晓杰(1990—),男,硕士研究生,主研方向为处理器功耗优化技术;安虹,教授、博士生导师;陈俊仕,博士研究生;迟孟贤、金旭,硕士研究生。
  • 基金项目:
    国家“863”计划项目(2012AA010901)。

Energy Efficiency Optimization Scheme for Multi-core Processor Under Power Consumption Constraint

QIU Xiaojie,AN Hong,CHEN Junshi,CHI Mengxian,JIN Xu   

  1. (School of Computer Science and Technology,University of Science and Technology of China,Hefei 230027,China)
  • Received:2016-04-15 Online:2017-04-15 Published:2017-04-14

摘要: 将处理器功耗控制在预算以下有助于降低散热成本和提升系统稳定性,但现有功耗优化方案大多依赖线下分析得到的先验知识,影响实用性,而集中式搜索最优策略的算法也存在复杂度过高的问题。为此,提出功耗优化方案PPCM。利用动态电压频率调整(DVFS)技术控制CPU功耗在预算内以提高处理器能效。同时,将功耗控制和功耗分配解耦合以提高灵活性。采用动态调整的线性模型估计功耗,通过反馈控制技术对其进行调节。以计算访存比为指标在应用间分配功耗,并考虑多线程应用特征进行线程间功耗分配。实验结果表明,PPCM比Priority算法速度平均提高10.7%,能耗平均降低5.1%,能量-延迟积平均降低14.3%。与PCMCA算法相比,其速度平均提高4.5%,能量-延迟积平均降低5.0%。

关键词: 功耗控制, 功耗分配, 能效优化, 动态电压频率调整, 计算访存比, 线程关键度

Abstract: Keeping processor power consumption under budget can reduce the cost of cooling and improve the system reliability.Most existing energy efficiency optimization schemes are profile-based offline schemes,which may reduce practicality.Furthermore,centralized algorithms which exhaustively search for the optimal configuration may be too complex.In this paper,a power consumption optimization scheme PPCM is proposed,which utilizes Dynamic Voltage and Frequency Scaling(DVFS) to control CPU power consumption under buidget and improve energy efficiency.PPCM decouples power consumption control and power consumption allocation to improve flexibility.It uses a dynamic linear model to estimate power consumption and manages it based on feedback control technology.It uses the ratio of computation to memory access as an indicator to allocate power consumption among applications,and then considers the features of multithread application and allocates power consumption among threads.Experimental results show that PPCM outperforms Priority algorithm by 10.7% in speed in average,5.1% in energy saving in average and Energy-Delay Product(EDP) is reduced by 14.3% in average.It is superior to PCMCA by 4.5% in speed in average and 5.0% in EDP in average.

Key words: power consumption control, power consumption allocation, energy efficiency optimization, Dynamic Voltage and Frequency Scaling(DVFS), ratio of computation to memory access, thread criticality

中图分类号: