Comparison of load-based and queue-based active queue management algorithms

A number of active queue management algorithms have been studied since Random Early Detection (RED) was first introduced in 1993. While analytical and experimental studies have debated whether dropping/marking should be based on average or instantaneous queue length or, alternatively, based on input and output rates (or queue length slope), the merits and drawbacks of the proposed algorithms, and the effect of load-based versus queue-based control have not been adequately examined. In particular, only RED has been tested in realistic configurations and in terms of user metrics, such as response times and average delays. In this paper, we examine active queue management (AQM) that uses both load and queuing delay to determine its packet drop/mark probabilities. This class of algorithms, which we call load/delay controllers (LDC), has the advantage of controlling the queuing delay as well as accurately anticipating incipient congestion. We compare LDC to a number of well-known active queue management algorithms including RED, BLUE, FRED, SRED, and REM in configurations with multiple bottlenecks, round trip times and bursty Web traffic. We evaluate each algorithm in terms of Web response time, delay, packet loss, and throughput, in addition to examining algorithm complexity and ease of configuration. Our results demonstrate that load information, along with queue length, can aid in making more accurate packet drop/mark decisions that reduce the Web response time.