We consider a queue fed by a large number, say n, of on-off sources with generally distributed on-and off-times. The queueing resources are scaled by n: the buffer is B=nb and link rate is C=nc.The model is versatile: it allows us to model both long range dependent traffic (by using heavy-tailed distributed on-periods) and short range dependent traffic (by using light-tailed on-periods).A crucial performance metric in this model is the steady-state buffer overflow probability.This overflow probability decays exponentially in the number of sources n. Therefore, if thenumber of sources grows large, naive simulation is too time-consuming, and we have to use fastsimulation techniques instead. Due to the exponential decay (in n), importance sampling with an exponential change of measureessentially goes through, irrespective of the on-times being heavy-tailed or light-tailed. Anasymptotically optimal change of measure is found by using large deviations arguments. Notably,the change of measure is not constant during the simulation run, which is essentially differentfrom many other studies (usually relying on large buffer asymptotics).We provide numerical examples to show that the resulting importance sampling procedure indeedimproves considerably over naive simulation. We present some accelerations. Finally, we give shortcomments on the influence of the shape of the distributions on the loss probability, and wedescribe the limitations of our technique.
long-range dependence importance sampling queueing theory large deviations asymptotics buffer overflow heavy-tailed random variables