Elephant Flow Modeling Based On SVM And Ensemble Learning

The growth in Internet penetration rate,network scale,and network traffic volume increase the network complexity,leading to a higher network bandwidth consumption,a lower network efficiency,and a poor user experience.Accordingly,it is significant to effectively alleviate the network traffic congestion to improve the network performance and resource usage.As elephant flows have large flow size,consume most of the network bandwidth,multiple elephant flows sharing a same link can congest the link,leading to a longer network delay that degrades network performance.Therefore,analyzing and modeling elephant flows and detecting them in their early stage early are the major concern of current research and this work.Many proposed approaches in current research use a threshold-based model to abstract elephant flows for their detection and prediction.Although some approaches use conventional classification algorithms for elephant flow modeling and analyzing,they typically do not fully consider the issue that elephant flows often have small quantities comparing with mice flows,leading to a low detection accuracy.This work addresses it by applying Machine Learning algorithms to model elephant flows,comparing the effectiveness of the models,and evaluating the generalization accuracy of the models based on multiple data sets.Particularly,this work firstly establishes a network data exporter,captures the traffic trace of Guilin University of Aerospace technology,and analyzes the characteristics of the trace.Next,this work selects the cumulative flow size and mean packet inter-arrival time of the top-k packets as the feature input for elephant flow modeling.Then,two types of elephant flow models based on Support Vector Machine(SVM)and Easy Ensemble are established.Regarding model evaluation,such two models are estimated using performance metrics over multiple data sets.Since a detection period should be considered in practice,we finally evaluate the performance of the proposed models given various detection periods,and proposed a scheme for optimizing the parameters of the original model and a scheme for first resetting the duration of the network data stream and then rebuilding the model.These solutions make it possible to detect elephant traffic using this type of model in a short detection time.The results show that the use of the model in this study can accurately detect about 71% of the elephant flow in the same data center type data set within 4.03 seconds.At the same time,the accuracy of the model detection can continue to be improved through the improved solution provided.In the modeling,we found that improving model detection accuracy while reducing detection time can be achieved by setting an appropriate network data flow duration.