Research On TCP Phases Estimation Based On LSTM Recurrent Neural Network

Network has penetrated into various fields such as social economy,culture and science,the performance of network affects the personal life directly or indirectly.A significant amount of today' s Internet traffic and network services,is carried by the TCP transport protocol.Therefore,the performance of network that users can perceive depends on the performance of TCP.TCP performance analysis is usually considered from the overall perspective of TCP data transmission such as throughput,delay or packet loss rate,but it cannot characterize more detailed features of the TCP data trans-mission process.And the TCP phases can reflect the performance of TCP from the data packet level,so as to analyze the factors affecting the TCP performance in more detail.This paper focuses on the TCP phases,mainly including building a TCP phase data set and using different algorithms to estimate the TCP phases.1)There is no TCP phases data set in the existing research.We study the TCP protocol,and modify the Linux kernel network stack to add the ssthresh and cwnd to the TCP header options.Then we build up a data collection platform to collect TCP data in multiple network scenarios,and use the added ssthresh and cwnd labeling TCP phases to get a TCP phase data set.2)We propose a rule-based TCP phases estimation algorithm to estimate the TCP phases.Firstly,we analyze the performance of TCP data transmission in different TCP phases.And then we use the ipmroved Self-Clocking algorithm to estimate the RTT and then estimate cwnd to design rules estimating the TCP phases.We use an over-lapped window smoothing technique to smooth the effect of the cwnd estimation error to improve the estimation accuracy of the TCP phases..3)We propose a TCP phases algorithm based on decision tree to solve the prob-lems that the estimation accuracy of the slow start phase is not high and the threshold parameters neeed to be manually set when designing rules.Then we obtain the TCP phases feature set by processing the TCP phases data set through feature engineering.Finally,we train and learn the TCP phases feature set to construct a decision tree model to estimate the TCP phases.Experiments show that the decision tree algorithm can improve the accuracy of the slow start phase.4)In order to further improve the estimation accuracy of the slow start phase,we propose a TCP phases estimation algorithm based on LSTM feature extraction.Firstly,the LSTM network is used to learning the features timing-related of the TCP data.Then the obtained features and some features obtained by the feature engineering are com-bined to form a new TCP phase feature set.Finally,we use a decision tree algorithm training the new TCP phases feature set to estimate TCP phases.The experimental re-sults show that the features extracted by the LSTM neural network can greatly improve the estimation accuracy of the slow start state.