Uniform Versus Priority Dropping for Layered Video

Wednesday, November 3, 1999 - 17:30
TH 331
Lee Breslau AT&T Research
Network routers drop packets on congested links during periods of overload. The decision about which packets to drop can treat all packets uniformly, or it can assign differential dropping priorities to different packets. Priority dropping offers the opportunity to increase total network utility by decreasing the dropping probability of more valuable packets while shifting loss to less valuable packets. However, it has been argued that the traditional uniform dropping used in the Internet, which gives all packets an equal probability of being dropped during congestion, should be retained because it provides applications with better incentives to constrain their usage. In this talk, I will discuss the relative merits of uniform and priority dropping. Using both simulation and analysis we modeled layered video applications under the two dropping regimes, comparing both their performance and incentive properties. Contrary to our previous intuition, we found that the performance benefit of priority dropping is smaller than expected, while the incentive properties of uniform dropping are worse than previously believed.

Lee Breslau received a PhD from the University of Southern California in 1994. Between 1994 and 1999 he was a member of the Computer Science Lab at the Xerox Palo Alto Research Center. Earlier this year he joined AT&T Research in Menlo Park, CA. His recent research has focussed on measurement-based admission control for real-time service, interactions between network control mechanisms and application performance, and web caching.