A framework for media-streaming over packet-switched networks
Peter Scott, Mahesh Venkatraman · 1999
Development of advanced video codecs and the wide-spread deployment of packet switched networks have facilitated delivery of multi-media content in a cost-effective manner. In our work, we develop mechanisms that will help increase the efficiency of this process. A video encoder and decoder work in a loosely coupled manner. The packetized data from encoder is transmitted to decoder through a series of packet-switches. For the encoder-decoder pair to function properly, data packet stream should admit characterization. Current packet switched networks (IP networks) cannot be well characterized, since they are build around a “best-effort” paradigm. The other extreme in supporting these codecs are ATM-type real-time networks which do not take into account the loosely coupled nature of the encoder-decoder pair. In this thesis we identify a better paradigm for supporting video streams. This is based on the concept of controlled-load traffic, where the amount of traffic on a link is limited. This is done by using an admission control algorithm. Traditional admission control algorithms expect sources to tightly characterize their streams making them less practical. In our framework, we use a measurement based admission control mechanism with loose traffic stream characterization. We chose a short-time exponential characterization to model the sources. Sources are characterized based on the backlog they generate at a constant rate server. An exponential bound on this backlog is used as the source model. Based on this model, the backlogs in a First-Come-First-Serve node and a multi-class Rate-Proportional-Processor-Sharing node are modeled. These characterizations are used in the admission control algorithms. In order to dynamically model the streams, a measurement based parameter estimation procedure is developed based on least-squares estimation. Using this measurement based admission control algorithm, a traffic management framework is developed which is ideal for streaming video sources. End-to-end characterization of data flow is performed by simple superposition of individual node characterization. This simplifies the traffic management process and thus makes it more robust. The developed algorithms are evaluated and validated using discrete event simulations. MPEG-I compressed video streams are used in the simulation studies. High degree of utilization is achieved in the network.