Modelling computer network traffic using wavelets and time series analysis

Modelling of network traffic is a notoriously difficult problem. This is primarily due to the ever-increasing complexity of network traffic and the different ways in which a network may be excited by user activity. The ongoing development of new network applications, protocols, and usage profiles further necessitate the need for models which are able to adapt to the specific networks in which they are deployed. These considerations have in large part driven the evolution of statistical profiles of network traffic from simple Poisson processes to non-Gaussian models that incorporate traffic burstiness, non-stationarity, self-similarity, long-range dependence (LRD) and multi-fractality. The need for ever more sophisticated network traffic models has led to the specification of a myriad of traffic models since. Many of these are listed in [91, 14]. In networks comprised of IoT devices much of the traffic is generated by devices which function autonomously and in a more deterministic fashion. Thus in this dissertation the activity of building time series models for IoT network traffic is undertaken. In the work that follows a broad review of the historical development of network traffic modelling is presented tracing a path that leads to the use of time series analysis for the said task. An introduction to time series analysis is provided in order to facilitate the theoretical discussion regarding the feasibility and suitability of time series analysis techniques for modelling network traffic. The theory is then followed by a summary of the techniques and methodology that might be followed to detect, remove and/or model the typical characteristics associated with network traffic such as linear trends, cyclic trends, periodicity, fractality, and long range dependence. A set of experiments is conducted in order determine the effect of fractality on the estimation of AR and MA components of a time series model. A comparison of various Hurst estimation techniques is also performed on synthetically generated data. The wavelet-based Abry-Veitch Hurst estimator is found to perform consistly well with respect to its competitors, and the subsequent removal of fractality via fractional differencing is found to provide a substantial improvement on the estimation of time series model parameters.

Grantor

Department of Electrical Engineering

Year dc:date.issued

2019

Author dc:creator

• Ntlangu, Mbulelo Brenwen

Advisor dc:contributor.advisor

• Baghai-Wadji, Alireza

• Engineering

Handle dc:identifier.uri

http://hdl.handle.net/11427/30146

OAI identifier oai:identifier

oai:open.uct.ac.za:11427/30146