Towards a theory of differentiated services

Architecting networks capable of providing scalable, efficient, and fair services to users with diverse QoS requirements is a challenging problem. The differentiated services framework has advanced a set of building blocks comprised of per-hop and access point behaviors. We provide a theoretical framework for reasoning about differentiated services networks, constrained to be implementable in IP networks. The control framework incorporates assumptions, albeit weak, about selfish user behavior and service provider behavior. This is necessitated by the essential role they play in influencing end-to end QoS, without which an effective evaluation of DiffServ architectures remains incomplete. We show that there is an intimate relationship between the properties exported by per-hop and edge control, and the "goodness" of the resource allocation and QoS attained in a noncooperative network environment. Our control framework-Scalar QoS Control-generalizes the per-hop and edge control achievable by setting a scalar value in packet headers, e.g., the TOS field of IP. We develop a theory of optimal classifiers and the properties they exhibit which facilitate end to-end QoS via the joint action of aggregate flow control per hop and per-flow control at the edge. We show the stability and efficiency properties of the overall network system when users are allowed to influence the choice of scalar values in the DS field at the edge, and service providers export costs to users commensurate with the QoS received.