Mining and control of network traffic by computational intelligence

Abstract

The structure and behavior of packet switched networks is difficult to model in a way comparable to many natural and artificial systems, Nonetheless, the Internet is an outstanding and challenging case because of its fast development, unparalleled heterogeneity and the inherent lack of measurement and monitoring mechanisms in its core conception. This thesis deals with applications of computational intelligence methods, with an emphasis on fuzzy techniques, to a number of current issues in measurement, analysis and control of traffic in the Internet. We design and analyze novel methods and address the hardware implementation of some of them.

First, we address mining of network traffic time series and traffic flow measurements. We develop a method for time series prediction by means of interpretable self-tuning fuzzy inference systems combined with a nonparametric residual variance estimation technique. The advantages of this method are illustrated through a number of time series benchmarks as well as a comprehensive set of traffic time series. Also, a method for the analysis and summarization of network traffic flow measurements by means of fuzzy linguistic summaries is developed. The method is shown to be fast and provide insightful and concise summaries.

Then, we focus on Internet traffic control. Methods for both end-to-end congestion control and traffic control at the IP layer are developed. As for end-to-end control, an extended TCP-like window based congestion control scheme with fuzzy inference is proposed and evaluated on realistic simulated scenarios as well as emulated and production networks. Its advantages are illustrated by comparison against traditional alternatives. As for IP layer control, fuzzy controlers are designed for active queue management and an in depth evaluation is performed on realistic simulation and emulation scenarios. These controlers are compared against traditional alternatives and their advantages are shown.

Finally, the hardware implementation of some of the methods developed in this thesis is studied. We propose an open FPGA-based platform and a companion methodology for developing fuzzy components for complex digital systems, specially in modern router architectures. The hardware implementation of fuzzy inference systems analyzed are shown to satisfy operational requirements of current and foreseeable future high-end routing hardware in terms of both inference speed and resource consumption.