Internet applications and users have diverse Quality-of-Service (QoS) requirements, making the current common-service-for-all paradigm inadequate and limiting. 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.
Our research has focused on relative service differentiation. This is a scalable and simple network architecture, because it does not require per-flow state in the routers, admission control, signalling, static routing, or resource reservations. In that context, we developed the Proportional Differentiated Services (PDS) architecture (IEEE Network, September 2000). 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. A central issue in this architecture is the design of forwarding algorithms (packet scheduling and buffer management) for proportional differentiation between classes.
We first considered the problem of delay differentiation and packet scheduling. The Proportional Delay Differentiation (PDD) model (ACM SIGCOMM 1999) requires a given ratio between the average delays of different classes. We designed three scheduling algorithms for the PDD model. 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. These three schedulers are evaluated in a more recent study (accepted for publication at the IEEE/ACM Transactions on Networking).
We also considered the issue of loss differentiation, and the related packet dropping problem. The Proportional Loss Differentiation (PLD) model (IEEE IWQoS 2001) constrains the loss rate ratios between classes. We designed and evaluated two dropping algorithms for the PLD model. The two droppers, PLR(infinity) and PLR(M), differ in the interval over which the loss rates are measured. This difference causes a trade-off between the two droppers, in terms of 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 investigated this Dynamic Class Selection (DCS) framework (IEEE ICNP 2001) in the context of proportional delay differentiation. We examined whether an acceptable class exists for each user, and show the properties of the resulting distributed DCS equilibria. Simulations provided further insight in the dynamic behavior of DCS, and in the factors that determine the well-provisioning of a network.
Finally, we considered the class provisioning problem (Scalability and Traffic Control conference, Aug 2001). The network provider, in this case, has an estimate of the rate and average delay requirement of each traffic type in a link. The objective is to compute the required link capacity and the appropriate parameters of the PDD model that meet these requirements. We designed a methodology that meets this objective.
Constantinos Dovrolis, September 2001