Active queue management (AQM) refers to a family of packet-dropping mechanisms for FIFO queues that has been proposed to support end-to-end congestion control mechanisms in the Internet. We examine the performance of AQM mechanisms by varying two parameters: the queue size and the dropping function. AQM flavors considered include both the well known "Drop from Tail" (TD) and "RED," as well as the more recently proposed "Gentle RED" and a previously unrecognized mechanism we call "Gentle RED with instantaneous queue size", or GRED-I.

We use experimentation to analyze the four AQM mechanisms identified above. The major result observed is that from a TCP goodput and UDP loss standpoint, varying AQM parameters has a minor influence on performance. On the other hand, we observe an important sensitivity of AQM to traffic characteristics. We show that RED and GRED do not provide sufficient performance improvement compared to Tail Drop. In the case of consecutive losses, we show that RED can be harmful by dropping much longer bursts than Tail Drop.

We conclude the presentation by discussing the feasibility of Fair Queueing as an alternative to Active Queue Management. Measurement data collected on the Sprint operational IP backbone will be used to advocate the deployement of Fair Queueing.

About the Presenter
Christophe Diot works with the Sprint Advanced Technology Laboratory, where he is in charge of the IP research group (www.sprintlabs.com). Sprint ATL's charter is to analyze new mechanisms and protocols for the Internet. Current activities in the IP area address diffserv, RED, and PIM-SSM deployment. The IP group's major project is the passive monitoring of the Sprint IP backbone in order to study IP traffic characteristics.