Endpoint Congestion Management IETF 46 -- Washington DC November, 1999 Chair: Vern Paxson Reported by: Mark Allman (mallman@grc.nasa.gov) Vern Paxson presented an overview of the working group and noted that the purpose of the meeting was to get everyone on the same page. The goals of the WG are to outline a set of congestion control (CC) algorithms that transport protocols can use rather than developing their own, and to develop ways to unify CC across connections. Vern noted that the granularity across which we share CC information is a difficult problem and the IRTF is being encouraged to examine the problem. Further, Vern noted that determining how much capacity exists on the network path needs to be orthogonal to scheduling which connection is allowed to use that capacity. The WG's scope is currently limited to protocols that have loss feedback. A comment was made that the WG is not looking at congestion control for the ACK stream or for streaming media such as video. Vern noted that we are doing what we understand, and encouraging the IRTF to investigate what we do not. Next, Hari Balakrishnan presented an overview of draft-balakrishnan-cm-01.txt. Hari noted that the Congestion Manager (CM) now includes an API hook for setting the "macro-flow" (connections which have agreed to share network capacity). A question was raised about whether the API was rich enough to allow policy servers to inject rules about certain flows. Hari noted this as something to think about. A concern was raised because application designers need to worry about all these messy congestion control details. Hari noted that using the CM API is much easier than implementing the CC algorithms in the application. In response to another question, Hari noted that the current version of the CM is window based (as opposed to rate based). A point was raised that we may need to have a "no network" message, in addition to the "no loss" messages currently included in the draft. This thought was extended by Aaron Falk to include communication between the link layer and the transport layer. Hari noted that this was a nice thought, but is researchy and should be put off until later. Joe Touch voiced the opinion that some of the numbers in the API should include scale factors with them, or there is a chance that we'll outgrow this system in the future. A short discussion of security indicated that if an application lies about loss we may get nasty denial of service attacks. Vern noted that this is a hard problems and we do not yet have a good answer, but also that because CM needs to only be deployed on one endpoint in order to work, it has a relatively easy deployment and upgrade path. Finally, Sally Floyd made a short presentation about draft-floyd-cong-00.txt, which is a general discussion of congestion control principles. Sally noted that traffic should be fair to TCP. Aaron Falk suggested that "fairness" is a researchy notion. Sally said that each proposal should be evaluated on its own and that "fair" does not have to be concretely nailed down. A question was raised about whether the CC principles applied to diffserv-like traffic. Sally indicated that they only apply to best effort traffic. Christian Huitema noted that since links to the home have very low bandwidth we are not in danger of congestion collapse from more aggressive TCPs and that the real solution to the problem is per flow scheduling. Sally disagreed, and the argument was made that even if this is the case, it's risky to bet the network's stability on this property. Hari Balakrishnan noted that we probably want both end-host mechanisms and router-based mechanisms.