In the best-effort Internet, congestion can cause severe degradations in the performance of both reliable data transfer flows and multimedia flows. These reliable data transfers are typically based on TCP, a responsive protocol. Responsive protocols are those that respond to congestion by reducing the rate at which they transmit data. This behavior is desirable because when all traffic is responsive, the traffic sources cooperate to ameliorate congestion by arriving at an aggregate operating point where the generated load matches the available capacity. In contrast, multimedia applications are typically based on unresponsive protocols, such as UDP, where the transmission rate is determined by the application, independent of network conditions. There exists a fundamental tension between these responsive and unresponsive protocols. When both traffic types are present during periods of congestion, unresponsive traffic maintains its load, forcing responsive traffic to reduce its load. Consequently, unresponsive traffic may benefit from this behavior and consume more than its fair share of the available bandwidth while responsive flows receive less than their fair share and suffer poor throughput. In the extreme, congestion collapse is possible. This offers a disincentive for using responsive protocols.

Recent proposals have attempted to address this problem by isolating responsive traffic from the effects of unresponsive traffic. They achieve this goal by identifying and severely constraining all unresponsive traffic. We take the position that some unresponsive traffic, specifically multimedia, is necessarily unresponsive. Maintaining the fidelity of the media stream places bounds on minimum levels of throughput and maximum tolerable latency. Since responsive protocols focus on reducing the transmission rate to match the available capacity independent of application-level concerns, these bounds may be difficult or impossible to meet with responsive protocols. As such, we argue that in addition to isolating responsive traffic, multimedia should not be unduly constrained and must also be isolated from the effects of other unresponsive traffic. In this dissertation we propose and evaluate a novel algorithm to allocate network bandwidth by allocating buffer space in a router's queue. This algorithm is called Class-Based Thresholds (CBT). We define a set of traffic classes: responsive (i.e., TCP), multimedia, and other. For each class a threshold is specified that limits the average queue occupancy by that class. In CBT, when a packet arrives at the router it is classified into one of these classes. Next, the packet is enqueued or discarded depending on that class's recent average queue occupancy relative to the class's threshold value. The ratio between the thresholds determines the ratio between the bandwidth available to each class on the outbound link.

CBT and other router queue management algorithms from the literature (FIFO, RED, and FRED) are implemented in a FreeBSD router and evaluated in a laboratory network using several combinations of TCP, multimedia, and generic unresponsive traffic. We show that CBT can be tuned to offer prescribed levels of performance for each traffic class. We present analysis that predicts network performance when using CBT based on initial configuration parameters, explain how this analysis can be reversed to derive optimal parameter settings given desired network performance, and empirically demonstrate the accuracy of this analysis. We present experimental data that contributes to our understanding of how existing queue management schemes perform, articulate relationships between parameters and performance metrics, and determine optimal parameter settings for each algorithm under various traffic mixes. We empirically demonstrate that CBT effectively isolates TCP while providing better-than-best-effort service for multimedia by comparing CBT's performance to the optimal performance for other algorithms. Finally, we show CBT provides better protection for TCP than RED and FIFO and better multimedia performance than RED, FIFO, and FRED.

Click here to get a PostScript copy of this dissertation (2.4 MB compressed), or here to get a PDF copy of this dissertation (2.6 MB uncompressed).