On the Relevance of Self-Similarity in Network Traffic Prediction

Self-similarity is an important characteristic of traffic in high speed networks which can not be captured by traditional traffic models. Traffic predictors based on non-traditional long-memory models are computationally more complex than traditional predictors based on short-memory models. Even online estimation of their parameters for actual traffic traces is not trivial work. Based on the observation that the Hurst parameter of real traffic traces rarely exceeds 0.85, which means that real traffic does not exhibit strong long-range dependence, and the fact that infinite history is not possible in practice, we propose to use a simple non-model-based minimum mean square error predictor. In this paper, we look at the problem of traffic prediction in the presence of self-similarity. We briefly describe a number of short-memory and long-memory stochastic traffic models and talk about non-model-based predictors, particularly minimum mean square error and its normalized version. Numerical results of our experimental comparison between the so-called fractional predictors and the simple minimum mean square error predictor show that this simple method can achieve accuracy within 5% of the best fractional predictor while it is much simpler than any model-based predictor and is easily used in an on-line fashion. I. I NTRODUCTION One of the key issues in measurement-based network control is to predict traffic in the next control time interval based on the online measurements of traffic characteristics. The goal is to forecast future traffic variations as precisely as possible, based on the measured traffic history. Traffic prediction requires accurate traffic models which can capture the statistical characteristics of actual traffic. Inaccurate models may overestimate or underestimate network traffic.