Traffic engineering models based on end-to-end loss probabilities and
delays do not scale well to fast backbone links.  In this paper, we
investigate the nature of congestion events in highly aggregated flows. An
examination of congestion events shows distinct phases of queue buildup,
packet dropping followed later by TCP reaction, and queue clearing.  
Evidence of the existence and characteristics of these discrete congestion
events is presented using active probe data gathered by the Surveyor
project.
When connections pass through periodic congestion, the aggregate offered
load to neighboring links rises and falls in cadence with the congestion
events.  We named this group of connections a "flock" and investigate the
implications. Flock formation, distinct from the prior notion of
synchronization,
can scale to larger numbers of connections depending on the diversity of
roundtrip times present.
Through simulation we illustrate conditions under which flocking occurs.
Unlike prior end-to-end studies, this paper takes the viewpoint of a small
number of hops along a long path. A model is presented that predicts
important characteristics of the congestion events including the quantity,
intensity, and duration.
Our results suggest that backbone congestion can be
modeled as an interaction between flocks of connections.  The model
effectively explains the prevalence of discrete congestion events
in fast links with high multiplexing factors
as well as the phases of congestion.