Speaker:  Dan Rubenstein
Affiliation:  U. Mass.
Date:  4/13/00
Time & Location:  4:15 PM, B17 Upson
Host:  Ken Birman
Title:  Fair Congestion Control for Large-Scale Continuous-Media Internet Sessions

The Internet is already being used extensively to transmit continuous media sessions consisting of audio, video, and/or text flows among numerous network users. Because these sessions can draw heavily upon network resources, it is important to design congestion control protocols that meet these sessions' requirements while "fairly" utilizing network bandwidth. This talk explores two techniques that can be used by congestion control protocols to help achieve these goals.

First, I will present how the use of data layering allows single-sender multicast sessions to achieve several desirable fairness properties. This examination is conducted using an extended definition of the notion of max-min fairness that accounts for layered multicast sessions. The practical limitations of coordinating receivers from the end-hosts is then explored, and I demonstrate that simple layered congestion control protocols can closely approximate the desired fair rates.

Next, I consider congestion control protocols that operate upon aggregates made up of multiple sessions. Such protocols are applicable in a variety of unicast and multicast  session scenarios in
which either a sender originates multiple sessions, or a receiver receives data from multiple sources. In both circumstances, a key piece of information that can be utilized by the congestion control
protocol is whether the sessions are bottlenecked by the same set of network resources. I will present novel techniques that make such a detection based on the comparison of the correlation statistics of packet loss and delay measurements for multiple flows, obtained at network end-systems. The techniques' effectiveness is demonstrated through analysis based on classical queueing models, simulation using realistic traffic, and experiments on the Internet.