Simultaneous and efficient simulation of highly dependable systems with different underlying distributions

Importance sampling is a well known technique that can be used for either variance reduction or obtaining performance estimates at multiple input parameter settings from a single simulation run ("what if" simulations). However, in queueing simulations, there is an essentially unique asymptotically efficient importance sampling distribution for estimating the probability y of certain rare events (e.g., buffer overflows). Furthermore, this unique distribution depends critically on the inputs of the model, thereby making it difficult to obtain good "what if" estimates from a single run. (An example of this is using a single run to estimate the mean time until buffer overflow at multiple arrival rates. ) In this paper, we show that a single importance sampling distribution can effectively be used for both variance reduction and "what if" simulation of certain rare events in models of highly dependable systems.