摘要： 网络排队时延对了解网络带宽利用率与分析拥塞级别具有重要意义，而传统时延测量技术对网络流量和往返时延预测的时效性差且准确性低，容易忽略突发的网络延时变化。结合交换机内部网络排队时延的细粒度特性和多变性，提出基于LSTM模型的多时间尺度融合预测方法。利用带内网络遥测技术获取并转换网络细粒度参数，为预测模型提供延时和利用率特征，构建基于长短期记忆网络（LSTM）的多时间尺度融合预测模型（LSTM-Merge），将不同采样尺度数据进行融合，并采用流式计算框架对网络排队时延进行预测。实验结果表明，与LSTM、SVR等预测模型相比，LSTM-Merge模型所得预测结果的均方根误差更小，3种时间尺度融合模型较其他数目时间尺度融合模型所得预测结果的实时性更好且准确性更高。

关键词: 长短期记忆网络融合模型, 网络排队时延, 时间序列预测, 流式计算, 机器学习

Abstract: Network queuing delay is of great significance for understanding network bandwidth utilization and analyzing congestion level.However,traditional delay measurement technology has poor timeliness and accuracy in predicting network traffic and round-trip delay,and it is easy to ignore sudden network delay changes.Combined with the fine-grained characteristics and variability of queuing delay in the internal network of switch,this paper proposes a multi-time scale fusion prediction method based on LSTM model.In-band network telemetry technology is used to obtain and transform fine-grained network parameters to provide delay and utilization characteristics for the prediction model.A multi-time-scale fusion prediction model(LSTM-Merge) based on Long Short-Term Memory(LSTM) network is constructed to fuse data of different sampling scales,and the flow calculation framework is used to predict the network queuing delay.Experimental results show that the Root Mean Square Error(RMSE) of the prediction results of the LSTM-Merge model is smaller than that of the LSTM,SVR and other models.Also,the real-time performance and accuracy of the prediction results of the three time scales fusion model are better than those of other scales.

Key words: Long Short-Term Memory(LSTM) network fusion model, network queuing delay, time series prediction, flow calculation, Machine Learning(ML)

中图分类号: 

• TP393.18