基于反馈机制的动态副本数量预测方法

Feedback Mechanism Based Prediction Method of Dynamic Replicas Number

针对视频点播集群服务中对单一视频访问数量进行预测时,基于传统的历史信息加权方法产生的动态副本数量误差较大、点播集群负载均衡较差的问题,分析了传统方法预测误差产生的主要原因——对单一视频历史访问数量趋势信息的考虑缺失,并提出了一种基于反馈机制的动态副本数量预测方法。该方法在传统的基于历史信息加权方法的基础上,考虑了历史访问数量的趋势信息。对实际系统数据的模拟结果显示,在反馈系数等于1时,基于反馈机制可以使动态副本数量的平均预测错误率下降29.6%,峰值错误预测率下降42.9%。

As for the prediction of the access number of a single video in Video-on-Demand（VoD） cluster service,it was reckoned by weighting the previous data.Such conventional solution,however,suffered the large error of dynamic replicas number and the imbalance of VoD cluster load.It was analyzed that the main reason causing large error of dynamic replicas number in conventional solution was the lack of consideration of the single video access trend.Thereby,in view of the video access trend as well as weighting the previous data,a new prediction method of dynamic replicas number was proposed,based on the feedback of the video access trend.The results in simulation of a real system trace show that,using the new prediction method based on feedback mechanism,in the case of feedback factor equals to 1,the average prediction error of dynamic replicas number decreases by 29.6%,and the peak prediction error decreases by 42.9%.