Efficient simulation of buffer overflow probabilities in jackson networks with feedback

Abstract

Consider a Jackson network that allows feedback and that has a single server at each queue. The queues in this network are classified as a single ‘target’ queue and the remaining ‘feeder’ queues. In this setting we develop the large deviations limit and an asymptotically efficient importance sampling estimator for the probability that the target queue overflows during its busy period, under some regularity conditions on the feeder queue-length distribution at the initiation of the target queue busy period. This importance sampling distribution is obtained as a solution to a non-linear program. We especially focus on the case where the feeder queues, at the initiation of the target queue busy period, have the steady state distribution corresponding to these instants. In this setting, we explicitly identify the importance sampling distribution when the feeder queue service rates exceed a specified threshold. We also relate our work to the existing large deviations literature to develop a perspective on successes and limitations of our results.