Analysis of self-similar traffic parameters for network performance improvement with real-time discrete wavelet transform

The main purpose of the present work is to estimate the Hurst parameter in real-time as a measure of network traffic self-similarity. There are many possible solutions for Hurst parameter estimation, but in this research the discrete wavelet transform approach has been used. The main reason is the possibility to calculate discrete wavelet transform in real-time for every input sample separately. There is no need to accumulate data and calculation can be made right away, which means that transformed data will be available faster for Hurst parameter estimation and processing load shall be distributed more evenly. The discrete wavelet transform provides other means for traffic as well, such as classification, anomaly detection, load prediction and so on. The Hurst parameter estimator algorithm has been proposed as well with algorithm diagram for software implementation. This algorithm has been implemented and its performance was tested and compared to regular discrete wavelet transform algorithm performance. The results show that the algorithm for estimating Hurst parameter is capable of performing estimation in real-time.

I. Introduction

At the present moment it is widely assumed that modern network traffic has self-similarity properties, which means the network traffic flow has similar behavior at different timescales (seconds, minutes, hours). The more scales keep this behavior, the higher is self-similarity of the traffic. This has been confirmed a long time ago by [1], where authors have conducted traffic measurements in different Bellcore laboratory local networks during time interval from 1989 and 1992. This work describes how data packets have been collected over short time intervals and probability distributions have been obtained. Such distributions were studied at lower scales by decreasing packet count time interval by 10 times for random distribution subinterval.