Modeling and Performance Analysis of Networks with Flow Transformations

Data flows in modern networks have exhibited increasingly complex prospect along with the evolving network infrastructures and emerging diverse applications. Performance evaluation must adequately consider this trend. The tractable queueing network as a traditional theoretical tool for performance evaluation can not satisfy the new requirement anymore. One reason is that the important assumption of queueing theory, i.e. Poisson arrivals, does not further hold accurate for many networks, within which the flows are more bursty, self-similar, or long-range dependent. Many alternative methodologies to the classical queueing theory appeared. Network calculus, over two decades after Cruz’s pioneer work in 1991, has established itself as one promising theoretical tool for assessing this kind of networks. It can deal with problems that are fundamentally hard for queueing theory based on the fact that it works with (probabilistic) bounds rather than striving for exact solutions. The complexity of network flows attributes itself not only to the upgrading infrastructure and enlarging number of users but also to the operations onto the flows caused by various algorithms, protocols, services, and even network topologies. The flows might be altered due to the operations along the path from source to destination, e.g., in a lossy network, being transcoded, randomly routed, or somehow processed for certain purpose like improving energy efficiency. This is called flow transformation. It is a great challenge for the existing queueing methodologies, including network calculus. One reason is, the basic assumption that the system is lossless when defining the performance metric delay, does not hold anymore. The other is, usually these flow transformations are random. This thesis addresses an extension of the network calculus to deal with the random flow transformations and provides the performance evaluation. The thesis mainly comprises of three inter-connected parts. The first is to develop the so-called stochastic data scaling elements to model the flow