计算机工程

• 体系结构与软件技术 • 上一篇    下一篇

基于RAPL的机群系统功耗限额控制

刘嵩,刘轶,杨海龙,周彧聪   

  1. (北京航空航天大学 计算机学院,北京 100191)
  • 收稿日期:2016-05-20 出版日期:2017-05-15 发布日期:2017-05-15
  • 作者简介:刘嵩(1989—),男,硕士,主研方向为计算机系统结构、分布式系统、功耗控制技术;刘轶,教授、博士;杨海龙,讲师、博士;周彧聪,硕士。
  • 基金项目:
    国家“863”计划重大项目(2012AA01A302)。

Power Capping Control for Cluster System Based on RAPL

LIU Song,LIU Yi,YANG Hailong,ZHOU Yucong   

  1. (School of Computer Science and Engineering,Beihang University,Beijing 100191,China)
  • Received:2016-05-20 Online:2017-05-15 Published:2017-05-15

摘要: 功耗管控是高性能计算系统和分布式数据中心管理的热点问题。当机房供电受限时需要对机群系统的功耗上限进行控制,使有限的电力适应供电容量的动态变化。为此,设计并实现一个基于RAPL的功耗限额控制系统。建立机群系统功耗模型,利用RAPL对CPU功耗限额的控制能力并结合功耗差额测量方法,将机群系统功耗上限控制在设定限额内,在此基础上尽可能减少程序性能的损失。实验结果表明,在较小的性能损失下,该系统可有效降低峰值功耗并将其稳定在限额内。

关键词: 高性能计算, 分布式数据中心, 峰值功耗, 功耗限额, 差额测量, RAPL技术

Abstract: The management and control of power has already become a hot issue in the area of management of High Performance Computing(HPC) system and distributed data center.In order to adapt the limited power to the dynamic change of the power supply capacity when the supply of energy is limited in the computer room,it is necessary to control the upper power limit of cluster system.Aiming at this problem,this paper designs and realizes a power capping control system based on RAPL.By constructing the power model of cluster system,utilizing RAPL’s capability of controlling the power consumption limit of CPU and combining the method of measuring the difference of power,it sets the upper limit of energy consumption of cluster system within the previously set power cap.On this basis,it tries to reduce the losses of performance as much as possible.The result of experiment shows that this system can reduce the peak power effectively with slight performance and keep it below the power cap stably.

Key words: High Performance Computing(HPC), distributed data center, peak power, power capping, difference measurement, RAPL technology

中图分类号: