Research Of Network Traffic Self-Similarity And Prediction

The feature of network traffic is studied mostly based on time series analysis of statistics. Time series are impersonal records of historical behavior of the system studied, and time series analysis is to analyze these records in order to find the mutual statistical dependence between these records and get the structural features and behavioral rules of the system studied. Then we can predict the future behavior of the system based on the structural features and behavioral rules.Fractal theory is widely used in various fields, and its most obvious character is self-similarity. Self-similar series is long range dependent, and its autocorrelation function decays asymptotically in hyperbolic manner when the lag increases. Hurst parameter represents the intensity of self-similarity of time series. Ever since self-similarity was introduced into the description of network traffic, it has been taken into consideration in network traffic modeling and prediction. So, how to estimate Hurst parameter more accurately plays an important role in research of network traffic self-similarity. This paper compares the accuracy and stability of the methods for estimating Hurst parameter through an experiment and validates the reasons for self-similarity of network traffic through a simulation experiment.Network traffic prediction plays an important part in design of a new generation of network protocols, network management and diagnosis, design of high quality router and improvement of Quality of service (Qos). From the perspective of different time scales and different time series models, this paper models and predicts with the traffic trace data which are taken from the internet traffic archive and from real environment. This paper also compares the prediction performance of different time series models under different time scales. Based on the analysis of results of all the experiments, we draw the conclusion that time series models produce large prediction error under smaller time scales (millisecond and second) and performance well under larger time scale (minute).