Proportional differentiated services for the internet
Konstantinos Dovrolis, Parameswaran Ramanathan · 2000
It is widely agreed that the Internet architecture should offer some type of service differentiation, so that some traffic classes get better QoS than others. This dissertation focuses on relative service differentiation, which is a scalable and simple network architecture. We develop the Proportional Differentiated Services (PDS) model. In PDS, the differentiation between classes is controllable, allowing the network provider to adjust the QoS spacing between classes, and predictable, providing higher classes with better service than lower classes independent of the load conditions. We first consider the issue of delay differentiation, and the related packet scheduling problem. The proposed Proportional Delay Differentiation (PDD) model constrains the average delay ratios between classes. The Proportional Average Delay (PAD) scheduler meets the PDD model when it is feasible; PAD, however, is not predictable in short timescales. The Waiting Time Priority (WTP) scheduler approximates closely the PDD model even in short timescales, but only in heavy load conditions. The Hybrid Proportional Delay (HPD) scheduler combines the PAD and WTP features. We then consider the issue of loss differentiation, and the related packet dropping problem. The Proportional Loss Differentiation (PLD) model constrains the loss rate ratios between classes. We design and evaluate two dropping algorithms for the PLD model. The two droppers, PLR(infinity) and PLR(M), differ in terms of their implementation complexity, accuracy, and adaptability to varying class loads. Users with an absolute QoS requirement can dynamically search for the minimum class that meets that QoS. We investigate this Dynamic Class Selection (DCS) framework in the context of proportional delay differentiation. We examine whether an acceptable class exists for each user, and show the properties of the resulting distributed DCS equilibria. Simulations provide further insight in the dynamic behavior of DCS. Finally, we consider the class provisioning problem. The network provider, in this case, knows the rate and average delay requirement of each traffic type in a link. The objective is to compute the minimum required link capacity and the appropriate parameters of the PDD model that meet these requirements.