Large deviations for a class of Markov processes modelling communication networks
Abstract
In this paper, we prove a sample path large deviation principle (LDP) for a rescaled process n^-1Q_nt, where Q_t is a multi-dimensional birth and death process describing the evolution of a communication network. In this setting, Q_t is the join number of documents on the set of routes at time t. Documents to be transferred arrive on route r as a Poisson process with rate _r and are transferred at rate _r_r(x) where x represents the state of the network, _r^-1 is the mean size of documents on route r and _r(x) is the bandwidth allocated to route r. We describe a set of assumptions over the allocation under which the LDP holds. Since we want the «classical» allocatio- ns to verify these assumptions, the difficulty is to deal with weak properties- . For example, _r(x) is assumed to be continuous on the set _r=x:x_r>0 but may be discontinuous elsewhere. Several examples are provided including the max-min-fairness allocation, a classical one in the context of data networks. Since the main object to work with is the local rate function, a great care has been devoted to its expression and its properties. It is expressed as the solution of a convex program from which many useful properties are derived. We believe that this kind of expression allows numerical computations.