Review - A Binary Feedback Scheme for Congestion Avoidance

Review of the paper "A Binary Feedback Scheme for Congestion Avoidance
in Computer Networks"
by Ivan Hernandez

The paper shows a fair Congestion Avoidance solution to avoid a
network congestion collapse. The solution is simple and does not add
neither overhead traffic (as the source quench does) nor overhead
processing. Nevertheless, it requires feedback from the network. The
authors called its solution Binary Feedback, because they use a bit (0
or 1) to pass information to the transport layer.

The general idea is fairly simple, as a packet may flow over multiple
routers to its destination, one or more of these may be congested. A
congested router sets a "congestion indication" bit on the ip packet
flowing in the forward direction. Then, this feedback is passed to the
destination's Transport Protocol (for simplicity I will use TCP),
based in this information TCP may adjust the window size (WS).

The routers are in charge of trigger the congestion avoidance
mechanism, once they detect a potential congestion. The detection
takes place if the average queue length is greater than a threshold,
using the average allows the router to not react prematurely or to
transient changes.

On the other hand, the receiver of the ip packet has to do more work
in order to decide how to interpret the router's feedback. First, it
has to determine how often update the window size, the receiver waits
enough time for the new window size take effect, this wait will reduce
the oscillation of the window size. After the update to the window
size, the receiver kill keep getting some packets with the feedback
on, it is OK, the receiver will ignore them enough time for the update
WS take effect. Next, with all the ip packets received, if at least
50% of the ip packets has the the binary feedback set to 1, the
receiver will decrease the window size, otherwise, it will increase
the window size. The increases will be done adding 1 to the window
size and the decreases will be setting the window size to window size
times .85. The authors show how these values improve the fairness and
minimize the oscillation of the window size.

The paper is clear and nice explained, the authors support their ideas
with analytical results of a previous paper. The simulations were
performed on a limited number of configurations, and the graphs are
really rough. What it is interesting about this paper is the approach
to solve network congestion from the perspective of the routers,
because these layer 3 devices do not have any congestion control
built-in mechanism. In the last two papers it is clear the oscillation
window size issue, in this paper they just minimize the range of
oscillation.
Received on Mon Sep 25 2006 - 21:41:11 EDT