Rare events in a polling system: Rays and Spirals

Abstract

It's a situation everyone dreads. A road is down to one lane for repairs. Traffic is let through one way until the backlog clears and then traffic is let through the other way to clear that backlog and so on. When stuck in a very long queue it is inevitable to wonder how did I get into this mess? We study a polling model with a server having exponential service time with mean $1/\mu$ alternating between two queues, emptying one queue before switching to the other. Customers arrive at queue one according to a Poisson process with rate $\lambda_1$ and at queue two with rate $\lambda_2$. We discuss how we get at a rare event with a large number of customers in the system. In fact this can happen in two different ways depending on the parameters. In one case one queue simply explodes and runs away without emptying. We call this the ray case. In the other spiral case the queues are successively emptied but in a losing battle as the system zigzags to the rare event. This dichotomy extends to the steady state distribution and leads to quite different asymptotic behaviour in the two cases.