[Colloquium] Greenwald talk today 2:45pm at TTI

[Meridel Trimble]

TOYOTA TECHNOLOGICAL INSTITUTE TALK 

Speaker: Michael B. Greenwald
University of Pennsylvania

Speaker’s homepage: http://www.cis.upenn.edu/~mbgreen/

Time: 2:45pm
Date: Wednesday, March 31st
Place: TTI-C (1427 E. 60th St. – 2nd Floor)
Refreshments provided 

Title: AHBHA: Managing Congestion through Adaptive, Hop-By-Hop, Aggregation

Abstract: 
The Internet may be among the most rapidly adopted technologies in history. 
However, the speed of its adoption has enshrined what {\em is} at the expense 
of what {\em ought to be}. Congestion control is particularly problematic: 
although we no longer fear congestion collapse, existing congestion management 
schemes are both brittle and complex. Brittle, because tools such as RED are 
extremely sensitive to parameter settings. Complex, because many problems have 
been addressed individually through distinct mechanisms that may interact in 
unpredictable ways. Examples of such problems are utilization of large 
bandwidth-delay product links, protection against non-congestion-aware flows, 
defense against denial of service attacks, unfairness between competing flows, 
interactions between "herds of mice" and "elephants", as well as others. 

In this talk I describe three related projects. First, I describe the design 
of AHBHA, a new congestion control architecture, and present some preliminary 
results. AHBHA is based on hop-by-hop feedback, local flow aggregation, and 
segregation of congested and non-congested flows. AHBHA addresses a large 
number of problems with a single set of mechanisms. 

The design of AHBHA required knowledge of actual congestion patterns in the 
Internet. I describe {\tt cing}, a tool that can accurately capture delay 
distributions on individual links in remote corners of the network without 
requiring any extensions to the network's infrastructure. 

To handle the vast number of observations returned by {\tt cing}, I present a 
one-pass space-efficient algorithm that summarizes large quantities of 
streaming data, while meeting a pre-specified guarantee on the accuracy of the 
summary. This algorithm improves upon the worst-case space requirements of the 
best known previous algorithms by a factor of $\Omega(\log N)$. Perhaps more 
importantly, empirically, it seems to require constant space in almost all 
realistic settings.
 
If you have questions, or would like to meet the speaker, please contact 
Meridel at 4-9873/mtrimble at tti-c.org
 
For information on future TTI-C talks or events, please go to the TTI-C Events 
page: http://www.tti-c.org/events.shtml