Traffic Modeling And Its Burstiness Boundary Analysis Based On Network Calculus

More and more researches have proven that there exists in various kinds of network some special properties such as Self-similarity, Long Range Dependency(LRD) and Heavy-tailness. This phenomenon contradicts with classic theories. Thus, several new models of self-similar traffic have been proposed, among which a LFSN(Linear Fractional Stable Noise)-based one is the primary object of this dissertation.Comparing with other model, LFSN-based traffic model is supported by general Central Limit Law and latest theoretical researching results of ON/OFF model. It takes both LRD and heavy-tail into well consideration, at the same time, it is parsimonious in the number of parameters with each a specific meaning. It also includes the result of FBM(Fractional Brownian Motion)-based model. These make LFSN an ideal model selection.Performance analysis is one of the most important reason for traffic modeling. Network Calculus, especially Stochastic Network Calculus(SNC), provides valuable methodology to apply this analysis onto large scale and real network, and to construct applicable Hierarchy of Stochastic Quality of Service(QoS). That is the reason why this area has attracted more and more attention including this dissertation.Basic theories of stable distribution and process are discussed in the beginning, along with the specific definition of LFSN and main contents of SNC, which are most basic but important concept used by this paper.To establish basic experimental environment, this thesis has done a lot of works in simulation, estimation of both stable distribution and process, dependency structure measurement. Finnaly, stochastic burstiness boundary of self-similar traffic is conducted. In general, main results of this dissertation are as followsAn improved FFT-based possibility density approximation algorithm of stable distribution is proposed, which takes advantage of the properties of stable distribution and thus adaptively chooses cut-off frequency and sampling interval, as the same time, is more accurate by frequency expansion without including addition overload.A fast simulation method of general LFSN is researched. This method has no limitation in the form of LFSN, as well as more efficiency, which makes massive simulation possible.A new model of self-similar network traffic based on general LFSN is proposed.. Current models are based on different form of specific LFSN on the consideration of mostly mathematic convenience, not of the properties of traffic itself. The new one takes maximum generalization which contains no constrain on the parameters of LFSN.Measurement of the dependency structure of LFSN is studied. When doing goodness of fit test, much attention was paid on the distribution test, yet we address that the dependency structure is much more important for those stochastic process which are both LRD and non-Gaussian. Three type of metrics and its corresponding statistics method are discussed in this paper, includes NACF, nCoV and gCoD.A new experiment method is proposed with a fake LFSN sample which has no dependence structure and thus suitable for comparisons. We apply this method into the testing of our general LFSN simulation algorithm and prove it correct on dependency structure. At the same time, we do the same test on real traffic data and show that gCoD is more suitable for skewed process such as network traffic.Stochastic burstiness boundary of self-similar traffic model based on general LFSN model is derived. This work combines traffic modeling and SNC traffic model, and gets better results useful for further"stochastic QoS"research. The result here is not only more general, but also far more accurate than current one, which make it possible for establishing stochastic QoS hierarchy.A new massive simulation experiment is conducted based on LFSN fast simulation. This simulation can get stochastic burstiness boundary statistically and prove our result correct. The methodology here is creative and unique.