Active Queue Management and Scheduling Methods for Packet-Switched Networks

Abstract

To support the myriad of envisioned communication products of the future, there is a need to develop a network infrastructure that can provide larger bandwidth, with better control of quality of service (QoS). However, with increasing demand for applications running over packet networks, congestion at the intermediate nodes (e.g., routers and switches) can be a serious problem. Consequences include long delays, large delay variation and high packet loss rates. Different solutions requiring varying levels of modification to the currently used algorithms have been proposed both for responsive (e.g., TCP) and unresponsive (e.g., UDP) protocols. However, most of the solutions are either too complicated to implement in real life or not general enough to be applicable to an arbitrary network topology. In this thesis, we investigate two mechanisms - active queue management (AQM), and scheduling - that can improve QoS in the packet networks. AQM techniques attempt to prevent congestion and regulate the queue length by sending congestion signals (i.e., dropping and/or marking packets) in a proactive manner, which would eventually cause the senders to decrease their sending rates. We use an analytic model derived for TCP in the literature to develop an AQM scheme that not only controls the queue length at the intermediate nodes but also distributes the resources fairly between the users. We present two different schemes that have different levels of complexity and performance. We also propose a distributed networking scheme that improves the performance of our new AQM algorithms. Although AQM schemes work well with responsive protocols such as TCP, the performance degrades for unresponsive protocols since unresponsive protocols do not change their packet sending rate in response to the congestion notifications sent by the network. Scheduling algorithms can regulate both responsive and unresponsive flows and can also provide guarantees on some QoS parameters, such as latency and fairness. We propose a new scheduling algorithm that can guarantee a better latency and fairness bound than the previous schemes with a similar complexity. This new scheme uses different queues for each round to simplify the common sorting problem with the scheduling schemes. Also a new notion, virtual round, is used to time-stamp incoming packets.