Structural properties of proportional fairness: stability and insensitivity

TL;DR

This paper characterizes proportional fairness in network bandwidth allocation, proves its stability under general conditions, introduces a reversible variant with explicit stationary distribution, and compares its insensitivity properties to balanced fairness.

Contribution

It provides a novel Fenchel--Legendre transform characterization of proportional fairness, extends stability results to more general models, and introduces a reversible, insensitive modification.

Findings

Proportional fairness is stable under Markovian routing.

A reversible, insensitive variant of proportional fairness is proposed.

Stationary distributions of the modified and balanced fairness share large deviations properties.

Abstract

In this article we provide a novel characterization of the proportionally fair bandwidth allocation of network capacities, in terms of the Fenchel--Legendre transform of the network capacity region. We use this characterization to prove stability (i.e., ergodicity) of network dynamics under proportionally fair sharing, by exhibiting a suitable Lyapunov function. Our stability result extends previously known results to a more general model including Markovian users routing. In particular, it implies that the stability condition previously known under exponential service time distributions remains valid under so-called phase-type service time distributions. We then exhibit a modification of proportional fairness, which coincides with it in some asymptotic sense, is reversible (and thus insensitive), and has explicit stationary distribution. Finally we show that the stationary distributions under modified proportional fairness and balanced fairness, a sharing criterion proposed because of its insensitivity properties, admit the same large deviations characteristics. These results show that proportional fairness is an attractive bandwidth allocation criterion, combining the desirable properties of ease of implementation with performance and insensitivity.