Tool Support for Fast Simulation Techniques !

Over the last 10 to 15 years there has been an increased interest in (and need for) the use of fast simulation techniques for the evaluation of the dependability and performance of computer communication systems [1]. Indeed, impressive simulation speed-ups of several orders of magnitude have been attained in the derivation of estimates for failure probabilities and unavailability measures (dependability area) as well as for blocking probabilities (performance area) as small as 10 4 to 10 12, in case the employed models belong to a well-studied class. Although the number of classes of models that can be marked “wellstudied” increases at reasonable pace, there still remain simulation models that cannot be easily accelerated. Irrespective of the fast simulation technique employed, be it importance sampling, the RESTART method or any other variance reduction technique, its application generally requires intimate knowledge of both the technique itself and the system (or model) being evaluated. Unfortunately, this is too much asked for most performance engineers working at real dimensioning problems in a (tightly scheduled) design trajectory of a complex computer communication system. What is needed to satisfy performance engineers is a set of tools and techniques that can be applied with relative ease (black box approach) for a wide variety of models. This implies that the performance engineer should be able to specify, in a user-friendly way, both the model and the measure of interest, and that the tool should give support in the selection of an appropriate fast simulation technique to determine the specified measure. Once a particular fast simulation technique has been chosen, e.g., importance sampling, the tool should allow for the experimentation with biasing schemes and support the selection process, i.e., the tool should support the user in the construction of an efficient simulation. As a first example in this direction, the governor component in the tool UltraSAN allows the user to change at a high level of abstraction the sampling strategy (the failure biasing or forcing scheme) in the importance sampling-based simulation of dependability measures, without having to bother about the necessary changes of likelihood ratios in the estimators; these are automatically handled by the tool [4]. For the RESTART method, the automatic computation of thresholds based on a pre-simulation and structural model information in the Petri net tool TimeNET as proposed by Kelling [2] can be seen as another good example.