电子学报 ›› 2019, Vol. 47 ›› Issue (6): 1220-1229.DOI: 10.3969/j.issn.0372-2112.2019.06.006

• 学术论文 • 上一篇    下一篇

异步I/O连接算法的峰值功率建模

杨良怀, 戚加欣, 范玉雷, 张璐, 梅建萍   

  1. 浙江工业大学计算机学院, 浙江杭州 310023
  • 收稿日期:2018-01-23 修回日期:2018-12-03 出版日期:2019-06-25 发布日期:2019-06-25
  • 作者简介:杨良怀 男,1967年出生于浙江新昌.毕业于北京大学,获博士学位,浙江工业大学计算机学院教授,主要研究方向为数据科学与大数据技术.在相关领域国内外重要会议与期刊等发表学术论文80余篇.E-mail:yanglh@zjut.edu.cn;戚加欣 男,1993年出生于浙江嘉兴.浙江工业大学硕士研究生,研究方向为数据库系统.E-mail:qijiaxin758@gmail.com
  • 基金资助:
    浙江省基金项目(No.LY18C130012,No.LY16F020032);国家自然科学基金项目(No.61502420,No.61702456)

Peak Power Modeling for Join Algorithms with Asynchronous I/Os

YANG Liang-huai, QI Jia-xin, FAN Yu-lei, ZHANG Lu, MEI Jian-ping   

  1. School of Computer Science and Technology, Zhejiang University of Technology, Hangzhou, Zhejiang 310014, China
  • Received:2018-01-23 Revised:2018-12-03 Online:2019-06-25 Published:2019-06-25

摘要: 峰值功率是影响数据中心能效的一个重要因素.本文提出一种功率感知数据库系统中连接算法的峰值功率估算方法,非运行时峰值功率的估算的挑战在于没有运行时的系统信息作为模型的输入.为克服估算困难,提出使用CPU密集度作为CPU功耗指示量,理论上分析了异步I/O连接算法在峰值功率发生阶段的特性,通过模拟连接算法峰值功率发生阶段算法行为来估算该阶段最大CPU密集度,通过CPU密集度与CPU功率的内在联系建立异步I/O连接算法的峰值功率预测模型.实验对数据库系统中采用异步I/O机制的四个连接算法时模型准确性进行了验证,结果表明所提预测方法具有较好的预测准确性,平均相对误差低于7%.

关键词: 峰值功率, 功率建模, CPU密集度, 数据库能效

Abstract: Peak power is a critical factor on the power consumption of a data center.This paper proposes a peak power estimation method to predict the peak power to join operations in DBMS.The challenge of non-runtime peak power estimation lies in that there is no runtime system information to use for model construction.To overcome this issue,this paper uses CPU-boundedness as the proxy of CPU power consumption and analyzes the characteristics of the peak power occurring stage of join algorithms with async I/O in theory.By simulating the behavior of this stage,we estimate the maximal CPU-boundedness of join algorithms.By examining the relationship between the CPU-boundedness and CPU power,the peak power models of join algorithms under different CPU execution frequency are hence constructed.Experiments validate the effectiveness of the proposed models on three typical join algorithms in DBMS with async I/O techniques.Results showed that our proposed methods had good accuracy with the average relative error less than 7%.

Key words: peak power, power modeling, CPU-boundedness, DBMS energy efficiency

中图分类号: