It is generally recognized that aggregated network traffic is self similar and that employment of traditional Poisson generators in network simulations may give misleading results when assessing performance. Many methods have been proposed to produce self similar simulation input. Five generators capable of continuous generation of arrival times are evaluated in this thesis with respect to their ability to reproduce a desired level of self similarity.
Our study shows that the packet generator supplied with the widely used commercial network simulator Opnet Modeler perform by far the worst. The remaining four generators all perform well. The statistical tests detect no difference among these four generators for the most interesting levels of self similarity. However, results indicate that the generator based on multiplexing ON/OFF sources may perform slightly better than the other generators, provided that more than 100 ON/OFF sources can be used.
This thesis also includes a study on the effects of self similar traffic in network simulations. We have compared network performance between to different setups; one with self similar traffic and one with traditional Poisson traffic. Our study shows a significant increase in latency when self similar traffic is deployed. We have also shown that traffic with the same level of self similarity can give rise to very different performance effects, depending on the reasons for its self similar behavior. This implies that level of self similarity in itself is not adequate to fully describe the performance effects of network traffic.