Summary: A Binary Feedback Scheme for Congestion Avoidance

This paper describes something that appears to be a predecessor to ECN in IP.
It starts out by describing the difference between flow control and congestion
control.
Flow control: "Ensures that the logical link has sufficient buffers at the
destination"
  ==> Not overloading your peer

Congestion control: "Having the various logical links in the network cooperate
to avoid/recover from the congestion of the intermediate nodes"
  ==> Not overloading the network

The paper also makes clear the difference between congestion control and
avoidance. Congestion control detects when the network has entered the early
stages of congestion collapse and recovers the network, but not before some
packet loss has occurred. Congestion avoidance keeps the network operating
just on the cusp of collapse, but doesnot drive it into this state.

The basic idea here is that when a router believes itself to be congested, it
sets the "congested bit" on all packets on one of the congested flows. This
information is eventually communicated back to the sender via transport-level
acks, which causes it to slow down. This method does not require dropping
packets to indicate congestion, as is ordinarily done. The paper defines
route as congested as any time a queue at a router serving that route has a
long-run average length greater than or equal to one.

Finally, the paper does some measurement analysis of an implementation of
their idea.
Received on Tue Sep 26 2006 - 03:59:34 EDT