• 中国计算机学会会刊
  • 中国科技核心期刊
  • 中文核心期刊

计算机工程与科学

• 高性能计算 • 上一篇    下一篇

一种基于周期性特征的数据中心在线负载资源预测方法

梁毅,曾绍康,梁岩德,丁毅   

  1. (北京工业大学信息学部,北京 100124)
  • 收稿日期:2019-08-20 修回日期:2019-10-26 出版日期:2020-03-25 发布日期:2020-03-25

A periodical characteristic-based resource
prediction method for datacenter online services
 

LIANG Yi,ZENG Shao-kang,LIANG Yan-de,DING Yi   

  1. (Faculty of Information Technology,Beijing University of Technology,Beijing 100124,China)
     
     
  • Received:2019-08-20 Revised:2019-10-26 Online:2020-03-25 Published:2020-03-25

摘要:

以Web服务、流式计算为代表的在线负载是数据中心的主要负载之一。在线负载请求到达的波动性驱动其资源需求的动态变化。因此,快速、准确的在线负载资源预测是数据中心合理分配资源、保障负载执行效率的关键。然而,既有在线负载资源预测方法或无法进行长期准确的预测,或依赖于海量样本数据并具有较大的时间开销。为此,提出了一种基于请求周期性特征的在线负载资源预测方法PRP。PRP面向在线负载请求的周期性特征,采用自相关函数识别负载资源使用的变化周期;基于变化周期进行资源使用样本序列分割及资源使用子序列分类;最终基于分类子序列采用线性加权方法预测在线负载的资源需求。实验结果表明,PRP在预测准确度和时间开销方面有较大的提升。
 

关键词: 数据中心, 在线负载, 周期性, 资源预测

Abstract:

Online services, such as web service and stream processing, are the major workloads in modern datacenters. The variations in request arrival rates lead to the dynamic changes in the online services’ resource demands during their executions. Therefore, fast and accurate resource prediction of online services is essential to reasonable resource allocation in datacenters and guaranteeing the efficiency of service execution. However, existing resource prediction methods of online services can neither conduct the long-term accurate prediction, nor conduct the prediction with limited sample dataset and low time overhead. To this end, this paper proposes a resource prediction method of online services based on the periodic characteristic of requests, called PRP. Oriented to the periodic characteristics of online ser- vice requests, PRP adopts the autocorrelation function to recognize the changing period of the online services' resource usage. Then, based on the changing period, it divides the resource usage sample sequence and classifies the resource usage subsequences. Finally, based on the classified resource usage subsequences, it uses linear weighting to predict the resources of online services. Experimental results demonstrate that PRP outperforms the existing resource prediction methods in both the prediction accuracy and the computational efficiency.
 

Key words: data center, online service, periodicity, resource prediction