Modeling, Simulation and Performance Evaluation
The design of a system is closely related to the ability to meet a reliable statement about the cost-benefit ratio. Compared to a complex hardware realization, modeling the system in question offers a good alternative. The system performance can be determined with a suitable model. Both, mathematical as well as simulative modeling methods can be used for this purpose.
Usually, mathematical methods are extremely time-consuming due to the complexity of the system to be modeled . Current and future research in this area target at an automatic generation of models. This helps eliminating the disadvantageous high design efforts for mathematical methods. First models have already been designed with a model generator, that is fed with simple rules about the system functionality. If changes of system functionality are desired, minor modifications of the rules must simply be provided to establish a new model. Nevertheless, detailed knowledge of the applied mathematical methods is currently necessary, as well as of the rule design. Advanced automation is supposed to further simplify the model design.
Simulation as an alternative suffers from high simulation run times to obtain reliable results whenever stochastic events are involved. Stochastic events arise because usually not all external influences on the system are known. The simulation's run time can be reduced by only simulating the relevant subsystem instead of the entire system. The behavior of the remaining system is approximated by probability distributions. If simulation is very close to the hardware level, for instance using VHDL, then there arises a problem of interaction between VHDL and stochastics (confidence intervals, random generators with specific distribution functions, etc.). Solutions are subject of research.
Further problem is that mathematical and simulative methods mainly provide only mean values of the performance measures in question. However, worst case results are also important, for instance concerning latencies in NoCs. Quantiles or the network calculus are to provide help. Finally, a transient performance determination is necessary, for instance for evaluation of the dynamic reconfiguration: Performance measures at certain points in time or in certain time intervals are important, i.e during a reconfiguration phase. Not all modeling methods can cope with this and are therefore subject to further research.
