The Differentiated Services (DiffServ) architecture is a promising means of providing Quality of Service (QoS) in Internet. In DiffServ networks, three service classes, or Per-hop Behaviors (PHBs), have been defined: Expedited Forwarding (EF), Assured Forwarding (AF) and Best Effort (BE).

In this dissertation, the performance of DiffServ networks at the network level, such as end-to-end QoS, network stability, and fairness of bandwidth allocation over the entire network have been extensively investigated.

It has been shown in literature that the end-to-end delay of EF traffic can go to infinity even in an over-provisioned network. In this dissertation, a simple scalable aggregate scheduling scheme, called Youngest Serve First (YSF) algorithm is proposed. YSF is not only able to guarantee finite end-to-end delay, but also to keep a low scheduling complexity.

With respect to the Best Effort traffic, Random Exponential Marking (REM), an existing AQM scheme is studied under a new continuous time model, and its local stable condition is presented. Next, a novel virtual queue and rate based AQM scheme (VQR) is proposed, and its local stability condition has been presented. Then, a new AQM framework, Edge-based AQM (EAQM) is proposed. EAQM is easier to implement, and it achieves similar or better performance than traditional AQM schemes.

With respect to the Assured Forwarding, a network-assist packet marking (NPM) scheme has been proposed. It has been demonstrated that NPM can fairly distribute bandwidth among AF aggregates based on their Committed Information Rates (CIRs) in both single and multiple bottleneck link networks.