关键词: 负载均衡, 时间序列的预测, 分类回归树, K近邻, 动态调度策略

Abstract: To address the problems in the mobile platform based on XMPP protocol, a prediction method of the server load based on classification and regression tree and K-nearest neighbor machine learning algorithm was presented. The algorithm made up time series of load by gathering every node's load information comprehensively and analyzed the time series, to carry out prediction. And then, the server nodes were divided into three regions, different scheduling strategies were used in different regions. Simulation experiments and tests showed that compared with the weight round robin and least connection algorithm, this proposed prediction algorithm decreased connection response time by 25%, and increased the connection establishment by 1.3 times. Dynamic scheduling strategy made the communication server cluster has a greater network throughput, which has more robust adaptability for mobile platform.

Key words: load balancing, forecast of time series, classification and regression tree, K-nearest neighbor, dynamic scheduling strategy