Techniques in Internet Congestion Control

This dissertation develops and analyses techniques for the control of congestion on IP networks. The two dimensions of congestion control are explored: 1) Load Control: control of amount of traffic transmitted onto the network 2) Capacity Dimensioning: provisioning enough capacity to meet the anticipated load to avoid congestion. We begin the work on load control by examining the design of Active Queue Management (AQM) algorithms. We focus on AQMs that are based on an integrator rate control structure. Unlike some previous AQMs, which measure congestion by measuring backlog at the link, this structure is based on the measurement of packet arrival rate at the link. The new AQMs are able to control link utilisation, and therefore control and reduce the amount of queuing and delay in the network. The rate-based AQMs can be used to provide extremely low queuing delays for IP networks, and enable a multi-service best-effort network that can support real-time and non-real-time applications. In the first part of Chapter 3, this dissertation develops a new rate-based AQM and a performance evaluation is performed comparing its performance to some key existing AQM proposals. In the second part of Chapter 3, the deployment of the rate-based class of AQMs is investigated experimentally and analytically, and critical efficiency issues with TCP/IP inter-operation are uncovered and addressed. In Chapter 4, the implementation of rate-based AQMs in multi-queue systems is also investigated. Rate-based control is analysed and applied to a multi-class Differentiated Services (DiffServ) scheduler, and a Combined-Input-Output-Queued Switch. In Chapter 5, the capacity dimensioning aspect of congestion control is addressed by the development of analytical tools for the performance evaluation of a multi-service link. The analytical tools predict the delay performance of traffic given link capacity parameters, and are useful for determining the required capacity to meet Quality of Service (QoS) requirements, such as for a convergent access link with Voice-Over-IP and Data traffic. The analytical tool uses matrix geometric methods and an iterative technique to solve the underlying M/G/1 queuing problem.