Reducing the complexity of realistic large scale Internet simulations

Computer simulation is a powerful tool for performance evaluation of new network protocols or dimensioning of communication networks. We show that current processors allow us to drive realistic simulations for medium to large network topologies with more than 110,000 HTTP/TCP clients. We reduce the complexity of the simulations by using a smaller number of clients as compared to reality while increasing the activity of each client and thus keeping the traffic load approximately constant. When evaluating how far the number of clients can be reduced without introducing large errors, we are focusing on the following measures: accuracy of matching a specified (measured) traffic load for each link, coefficient of variation, Hurst parameter, end-to-end delay and loss probability. We show that the only parameter affected by the reduction is the loss rate. Our results show that a significant reduction of the required memory, a factor of 4-8, can be gained by reducing the number of clients and increasing the client activity. A side-effect is a gain in simulation speed of up to 30 %. Nevertheless, the reduction of the required memory makes it possible to perform realistic simulations of large networks topologies on computer systems with hard per process memory limitations.