Differentiated quality of service in packet-switching networks

The advent of new network applications with diverse traffic characteristics and quality of service constraints demands networks to provide differentiated Quality of Service (QoS). In a QoS-enabled network, the differentiated active queue management and the packet scheduling are two of the most important mechanisms to support differentiated QoS at a network node. This research addresses the design issues in differentiated active queue management schemes and packet scheduling disciplines. In the first half of this research, we proposed discrete-time analytical models for differentiated active queue management schemes. The models were used to derive various queueing performance metrics. We identified the dropping probability function as the control knob in differentiated QoS provision, and investigated the optimal selection of dropping probability function parameters. We demonstrated that, with appropriate selection of dropping probability function parameters, differentiated QoS can be provided to contending traffic classes. Additionally, we compared the optimal queueing performance of different types of dropping probability function, and the range of service differentiation achievable using differentiated active queue management schemes. We found that the stair-case dropping probability function had the best performance, followed by the piece-wise linear and exponential functions. We also found that differentiated active queue management schemes provide a wide range of throughput differentiation and a limited range of delay differentiation to contending traffic classes. In the second half of this research, we proposed a statistical bounding approach to find the upper-bound of loss probability in a Generalized Processor Sharing (GPS) server. Using the derived loss probability upper-bounds, we studied the minimum size of GPS buffer allocation to satisfy QoS constraints of each GPS traffic class. The numerical results showed that our statistical bounds reduce buffer allocation by up to 10-fold compared with that derived from deterministic bounding approach. Additionally, we found the existence of a critical weight in GPS system. The findings could help choosing appropriate GPS weights.