Methodology for parameterisation of large scale network simulations

Network providers are confronted with the optimisation or extension of existing networks. This leads to new challenges for simulation-based network dimensioning. The first challenge is the realistic simulation of the existing network, where topology information and traffic measurements have to be considered. The second challenge is to predict how changes of the existing network will affect its performance: link capacities, queue management algorithms etc. are subject to changes to adapt the network to changing traffic requirements. The best strategy for enhancing or extending the network under consideration can be found by comparing the resulting benefits and disadvantages between the existing network and new alternatives. The simulation of existing networks is an inverse problem: (i) the network description and some measurements are given from the network provider; (ii) the number and the behaviour of clients must be derived from the given parameters. Considering this problem was motivated by an industry project “ERNANI” funded by the “Deutsches Forschungsnetz” (DFN) and the German Telekom. The author proposes a new methodology to solve the inverse problem. First, an algorithm for the allocation of clients is proposed. This algorithm efficiently controls the average traffic intensity. Second, a method is developed for matching higher-order moments of the simulated traffic to measurements made in existing networks. Two higher moment parameters were selected, that are of major importance in network engineering: the coefficient of variation and the Hurst parameter. Realistic network simulations are characterised by high complexity because internal states of protocols must be stored for each connection. Therefore, memory requirements and simulation speed are major issues in such simulations. One solution for this problem is to reduce the number of clients by increasing the activity of each client in order to keep the traffic characteristics unchanged. To asses the applicability and performance of this solution, critical network parameters are estimated as a function of the number of clients. The parameters considered are: average link load, loss probability, coefficient of variation of the packet inter-arrival times, Hurst parameter and average end-to-end delay. It is shown in this work that the number of clients, as well as the required memory, could be reduced by a factor of 4 − 8 without significant impact on the studied parameters. Reducing the number of clients by a factor of 8 the simulation speed increased by approximately 33 %. This work represents a major step towards realistic modelling and simulation of existing networks. The simulation results based on the presented methodology are very promising. The successful increase of the simulation efficiency represents one step towards the realistic simulation of current and future multi-Gbit networks.