Research On Network Traffic Characteristic Analysis And Congestion Control

With the popularization of network application and increase of communication services, Internet is becoming more and more busy. It's very important for effective control and management polices to be implemented. For a long time, traffic modeling and analysis is based on Poisson distribution and Markov process. But recently measures of network traffic have shown the self-similar nature of network traffic. And this property can't be described by traditional model. Under the support of project of "self-similar network traffic research based on alpha stable process" granted by multimedia and network communication engineering laboratory, some research work has been done in the area of theory foundation of traffic modeling, network design and traffic control.Firstly, self-similar phenomenon in network traffic is depicted. Definition and properties of self-similar process are presented. A comprehensive introduction of research results in the area of traffic self-similar is given. All these form a basis for latter research work.Secondly, this paper proposes the use of alpha stable distribution in traffic modeling. Analysis based on the combination of alpha stable distribution and properties and network traffic property shows validity of application of alpha stable process to traffic modeling. Moreover, this point is verified through statistical analysis of real traffic data. A representative trace is selected and regrouped separately according to packet count in unit time and throughput in unit time. Then this trace is thoroughly analyzed according to validation theory of alpha stable distribution. The result shows that network traffic arrival process can be accurately described alpha stable process.Thirdly, impact of self-similarity on network performance is discussed. If the input is self-similar traffic, the tail of mean queue length in buffer decreases as Weibullian distribution or hyperbolic function. This is different from result concluded from Poisson model. These properties induce large queuing delay and high packet loss rate. And the deterioration of performance becomes more evident when there is high self-similarity. If a larger buffer is used, packet loss rate will be lower to some extent. But queuing delay will increase dramatically. On contrary, if network bandwidth is enhanced, both the packet loss rate and queuing delay will decrease. Therefore, it's vital to determine appropriate buffer when network is planned.At last, a new congestion control scheme is proposed based on prediction of network traffic. It takes advantage of long range dependence to predict traffic level in the future period. Then the sending rate is adjusted according to the prediction result. Thus network resource is fully utilized and congestion can be avoided to some extent. In this paper, traffic modeling based on linear fractional stable noise process is given and prediction algorithm is presented. This algorithm can be applied to the proposed control scheme to improve additive increase multiplicative decrease rate control. Simulation is conducted to analyze the performance of the proposed scheme. The simulation results show that this scheme can improve throughput and decrease packet loss rate. Finally, the practicability of this scheme is analyzed.