Efficient simulation of network performance by importance sampling

The dissertation deals with simulation techniques for simulation of s ystems where the events of interests are rarely occurring relative to the tota number of events. A prohibitively long simulation time will be required to assess their occurrence probabilities without the app lication of special techniques. The focus of the work is on the impor tance sampling technique. The dissertation is aimed at simulation of the traffic handling performance and dependability of communication n etworks. However, since the underlying model is a continuous time dis crete state Markov model, the results of the dissertation are applica ble to a wider range of simulation problems. The results of the diss ertation are a significant step towards making simulation with import ance sampling a practical tool for performance evaluation of networks with strict quality of service requirements. This technique has unti l now only been applicable to simple systems where there is only one set of resources in the system or where one set of resources limits i ts performance. Based on the theoretical insight from one-dimensional systems a dynamic adaptive parameter biasing strategy is devised whi ch enables simulation of balanced multi-dimensional systems. The di ssertation investigates different methods and application fields that have been treated separately. This specifically refers to the compar ison and methods of combining the two main rare event simulation tech niques Importance sampling and RESTART. Furthermore, the dissertatio n proposes heuristics to decide if the importance sampling works as i ntended and if trustworthy results are obtained. As an appendix to th e dissertation, co-authored with others, original results concerning importance sampling in trace driven MPEG simulations are also reporte d.