Gábor Lencse:

Event-driven discrete event simulation using computers is a powerful method in the performance analysis of communication systems. The simulation of large and complex systems requires a large amount of memory and computing power that is often available only on a supercomputer. Efforts were made to use clusters of workstations or multiprocessor systems instead of supercomputers, as this would be much more cost effective. The conventional synchronisation methods for parallel simulation (conservative, optimistic) (Fujimoto, 1990) produce same results as sequential simulation but they are unfortunately not applicable to all cases, or do not provide the desirable speedup. There is a third method called Statistical Synchronisation Method (SSM) (Pongor, 1990) which can only be used in its original form to determine the steady state behaviour of the systems.

Starting from SSM, I have developed SSM-T, the time driven version of SSM. SSM-T can also be used even if the state of the investigated system changes sometimes. The applicability criteria of the method and the achievable speed-up were determined. It was proven that results of the parallel simulation using SSM-T can approximate the results of the sequential simulation with arbitrary precision. In a case study, I achieved 1.91 or 1.86 speed-up by a two processor parallel simulation using SSM-T compared to the sequential simulation.

The resource requirements and the accuracy of the statistics collection methods for SSM-T were examined. The L₁ error criterion was used. Depending on the distribution, suggestions were given for the statistical collection method to be used.

A statistics exchange control algorithm was proposed that is based on prediction and synchronisation point deletion. The so-called penalty functions were introduced. They were used to give a mathematical criterion that can be a measure of the goodness of the different statistics exchange control algorithms. The expected value of the overall penalty has been analytically determined. According to the simulation results, we can get near-optimal results for a relatively wide range values of the parameters of the prediction, thus SSM-T is robust enough to tolerate some inaccuracy of the parameters of the prediction.

Seven data structures were examined by simulation and compared to find out which one should be used to implement the Future Event Set of an event driven discrete event simulator. Though balanced trees yielded satisfactory performance, heap and especially skip list was found to be significantly better on modern processor architectures.

For the fast performance analysis of communication system, a new method called Traffic Flow Analysis (TFA) was developed. TFA is combination of simulation and the numerical or analytical methods. The applicability criteria of the method were presented. An example implementation of the method was described. In a case study, the results of the TFA method were close to the results of the pure analytical method.