Abstract: It is readily apparent that support for quality-of-service "in the Internet" is required in order for distributed, interactive multimedia applications to be effective. There are currently two working groups within the IETF that are addressing the problems of realizing network-based quality- of-service: the integrated services working group and the differentiated services working group. The groups differ significantly in their philosophy and approach to realizing QoS, and widespread deployment of technology promulgated by these groups is still a long way off. However, within these groups, there is widespread agreement that routers throughout the network should, today, perform "active queue management" -- explicitly discard packets before queues become full and overflow. The is done to (1) force flows to react to early signs of congestion and thereby decrease the size of queues in the network and improve overall network performance, and (2) to protect the network against so- called unresponsive and non-TCP compatible flows.
This latter goal does not bode well for interactive multimedia applications. As the vast majority of interactive multimedia applications use UDP for data transport, they are considered (by definition) to be "aggressive" and may be subjected to particularly harsh treatment in routers. Thus until integrated or differentiated service are deployed and pervasive, the performance of multimedia applications potentially may get worse before it gets better.
To mitigate this situation, we are working on active queue management policies for routers that attempt to balance the concerns of congestion collapse avoidance and the requirements for continuous transmission for interactive UDP applications that react to signs of congestion. Specifically, we are experimenting with extensions to the Random Early Detection (RED) packet discard mechanism for providing better performance for UDP flows without sacrificing performance for TCP flows.
This talk will review some of the current proposals for active queue management and attempt to make the case for a "better-than-best- effort" service for UDP. I will present our proposed extensions to RED and wave my hands vigorously to argue that they are beneficial. (Thus, this will largely be a speculative, work-in-progress talk.)