Distributed Algorithms for Rate-Adaptive Media Streams

Abstract

Rate adaptive media streams bear similarities to both elastic (e.g., file transfer) and inelastic (e.g., voice) network traffic classes. As such a service model for streaming media should incorporate characteristics of both classes: guaranteed minimum rates (for minimally acceptable quality of service (QoS)) and an efficient mechanism for intelligently allocating excess capacity among the competing streams. In this paper we apply the system decomposition Kelly (Charging and rate control for elastic traffic. European Transactions on Communications 8:33–37, 1997) and distributed algorithms framework Kelly et al. (Rate control in communication networks: shadow prices, proportional fairness, and stability. J Oper Res Soc 49:237–252, 1998), developed by Frank Kelly et al. in the context of elastic traffic, to a service model for rate adaptive media streams. Our aim is to allocate excess capacity among competing streams so as to maximize a weighted client-average QoS, where each client QoS is defined as the time-average utility of the instantaneous received rate. We develop the corresponding system decomposition and distributed algorithms appropriate for rate adaptive media streams. Two distinct classes of distributed algorithms are derived: i) a distributed algorithm for dynamic adaptation, where the instantaneous subscription level of each stream is adjusted in response to network updates on route congestion levels, and ii) a distributed algorithm for static adaptation. The stationary point of the static adaptation algorithm is an admission policy that assigns each stream a fixed (in time) transmission rate as a function of the stream volume (the total number of bits associated with the media object). Our results, consistent with earlier work Weber and de Veciana (Rate adaptive multimedia streams: optimization and admisssion control. IEEE/ACM Trans Netw 13(6):1275–1288, December 2005a), confirm the intuitive idea that client average QoS is maximized by granting preferential treatment to small volume streams over large volume streams. Moreover, we argue that static adaptation is able to achieve a comparable QoS to that of dynamic adaptation without the communication overhead and fluctuations in service required of the latter. The caveat is that static adaptation policies work best for stationary traffic loads. Extensive simulation results demonstrate the algorithms to be feasible and efficient.