SUMMARYSince the seminal work by Kelly on distributed network resource allocation using the language of network utility maximisation (NUM) a decade ago, there have been extensive research efforts generalising and applying NUM to model, analyse and design various network protocols and architectures. Some of these works combine the distributed optimisation approach with stochastic network models to study NUM under network dynamics occurring at the session, packet and constraint levels. We survey these works by presenting the key questions, results and methodologies in this emerging theory of stochastic network utility maximisation, followed by discussion on related work and future research challenges.
INTRODUCTIONIn the 1997 paper in this journal [1] and the 1998 paper [2], Kelly et al. presented an innovative idea on network resource allocation that has led to many research activities since. An optimisation problem is formulated where the variables are the source rates constrained by link capacities and the objective function captures design goals:where the source rate vector x is the set of optimisation variables, one for each of the sources indexed by i, the {0, 1} routing matrix R and link capacity vector c are constants, and U i (·) is the utility function of source i. Because the above network utility maximisation (NUM) problem can be readily decomposed, distributed algorithms are developed where each of the links and sources controls its local variable, such as link price or source rate, based on local observables, such as link load or path price. By techniques such as Lyapunov function or the descent lemma, global or local asymptotic convergence towards the optimum can be * Correspondence to: Mung Chiang, Department of Electrical Engineering, Princeton University, NJ 08544, USA. E-mail: chiangm@princeton.edu proved for these distributed algorithms, for cases with or without propagation delay. A key insight is that the effects of network protocols can be understood as the trajectories of a controlled dynamic system. After a decade of work by many researchers, there is now a substantial set of theory, algorithms, applications and even commercialisation based on the NUM model of networks.First, NUM and its extensions can be used to model various resource allocation problems and network protocols. Utility functions may depend on rate, latency, jitter, energy, distortion, etc., may be coupled across users, and may be any non-decreasing function, although most papers assume smooth and concave utility functions. They can be constructed based on user behaviour model, operator cost model or traffic elasticity model. They can also shape the fairness of resource allocation: a maximiser of the α-fair utility functions satisfies the definition of α-fair allocation. Here α-fair utility function refers to a family of functions parameterised by α 0: Sometimes, network protocols are modelled by NUM via 'reverse-engineering': a given protocol, originally designed based on engineering intuitions, is shown to be implicitly so...