MIT Department of Electrical Engineering & Computer Science

Scalable Internet Network Engineering

Robert Morris
Harvard University

Monday, March 29, 1999
4:15 PM (refreshments 4:00)
Room NE43-518
EECS Special Seminar

Abstract

Anyone who has used the Internet can testify that its performance is often erratic: transfers that sometimes complete almost instantly can take many seconds at other times. Much of this delay is caused by the network discarding data when heavily loaded. Unfortunately, little more than this is generally known about the behavior of the Internet protocols under high load. This lack of understanding has led to an ad-hoc approach to the design and configuration of network components, particularly router buffering algorithms.

This talk presents two contributions. First, a framework for analyzing load on Internet links. The framework's metric for load is the number of competing connections, rather than the amount of bandwidth consumed; the metric for capacity is the total storage in the network, rather than the network's bandwidth capacity. These metrics allow predictive and explanatory analyses of load and packet loss. The second contribution is a pair of router queuing algorithms (FRED and FPQ). These algorithms adapt the amount of available buffering to the number of active connections, allowing them to provide good delay performance across a wide range of loads. Networks employing these algorithms will serve users with consistent and predictable delay even under heavy load.