Reversibility in Queueing Models

Abstract: In stochastic models for queues and their networks, random events evolve in time. A process for their backward evolution is referred to as a time reversed process. It is often greatly helpful to view a stochastic model from two different time directions. In particular, if some property is unchanged under time reversal, we may better understand that property. A concept of reversibility is invented for this invariance. Local balance for a stationary Markov chain has been used for a weaker version of the reversibility. However, it is still too strong for queueing applications. We are concerned with a continuous time Markov chain, but dose not assume it has the stationary distribution. We define reversibility in structure as an invariant property of a family of the set of models under certain operation. The member of this set is a pair of transition rate function and its supporting measure, and each set represents dynamics of queueing systems such as arrivals and departures. We use a permutation {\Gamma} of the family menmbers, that is, the sets themselves, to describe the change of the dynamics under time reversal. This reversibility is is called {\Gamma}-reversibility in structure. To apply these definitions, we introduce new classes of models, called reacting systems and self-reacting systems. Using those definitions and models, we give a unified view for queues and their networks which have reversibility in structure, and show how their stationary distributions can be obtained. They include symmetric service, batch movements and state dependent routing.