电子学报 ›› 2015, Vol. 43 ›› Issue (10): 1904-1910.DOI: 10.3969/j.issn.0372-2112.2015.10.004

• 学术论文 • 上一篇    下一篇

GPU集群能耗优化控制模型研究

王海峰1,2, 曹云鹏1,2   

  1. 1. 临沂大学信息学院, 山东临沂 276005;
    2. 山东省网络环境智能计算技术重点实验室临沂大学研究所, 山东临沂 276005
  • 收稿日期:2014-03-06 修回日期:2014-06-02 出版日期:2015-10-25
    • 通讯作者:
    • 王海峰
    • 作者简介:
    • 曹云鹏 男,1967年10月出生,山东沂南人.临沂大学副教授.1989年和2005年分别在南开大学和山东科技大学获得计算机应用专业的学士和硕士学位,主要研究方向为并行计算、智能控制和仿真.E-mail:lyucyp@163.com
    • 基金资助:
    • 山东省自主创新及成果转化专项 (No.2014ZZCX02702); 山东省自然科学基金 (No.ZR2013FL005); 临沂大学博士科研启动项目 (No.2014LYDXBS018)

Power Consumption Optimization Control Model of GPU Clusters

WANG Hai-feng1,2, CAO Yun-peng1,2   

  1. 1. Information School, LinYi University, Linyi, Shandong 276005, China;
    2. Institute of Linyi University of Shandong Provincial Key Laboratory of Network based Intelligent Computing, Linyi, Shandong 276005, China
  • Received:2014-03-06 Revised:2014-06-02 Online:2015-10-25 Published:2015-10-25
    • Supported by:
    • Independent Innovation and achievement Transformation Special Project of Shandong Province (No.2014ZZCX02702); Natural Science Foundation of Shandong Province,  China (No.ZR2013FL005); Doctoral Research Project of Linyi University (No.2014LYDXBS018)

摘要:

随着大数据技术的发展,GPU集群作为一种高效的并行系统被应用到大规模数据实时计算中.能量是实时计算时重要的资源,GPU集群的能耗优化及实时消减成为一个具有挑战性的问题.从集群全局角度引入模型预测控制策略,并建立闭环反馈机制的多输入多输出控制器.通过调整计算频率和改变活跃流多处理器来改变能耗状态,利用反馈和滚动优化机制完成对未来的控制预判,实现消减冗余能耗的目标.实验表明:控制模型的精度和节能效果优于基准模型,而且具有较好的稳定性,适合应用到大规模数据实时计算中.

关键词: 能耗控制, GPU集群, 能量消减, 模型预测

Abstract:

With the development of Big Data technology GPU cluster as a high efficiency parallel system applies into the Large-scale data computing field.Energy is a significant computation resource.So power consumption optimization control and capping in real-time becomes a challenge issue.The Model Prediction Control strategy is introduced and a Multi-Input Multi-Output controller is built by using a closed loop feedback principle from the whole cluster perspective.Power consumption status is changed by scaling frequency and adjusting active stream multi-processors.Then the feedback and the periodic optimization mechanisms can predict the control behaviors in the future control cycles.This achieves the goal that reduces redundancy energy.The results demonstrate that the proposed model has more accuracy and comsumes less energy than the others.And it has better control stability.So it has better adaptability and obvious advantage in the Large-scale data real-time computing.

Key words: power consumption control, graphic processing unit (GPU) clusters, power capping, model prediction control

中图分类号: