Binary Feedback Review

This paper discusses a scheme for congestion avoidance, which is
distinguished from congestion control, i.e. attempting to eliminate
congestion conditions. Congestion avoidance aims to operate the network at
a highly desirable point, which is at the knee of the response time
curve. Therefore
it aims to use the routers efficiently while achieving fairness across all
users. It is similar to the "Congestion Avoidance and Control" paper but it
requires a change in routers and the transport layer header to allow each
router along a route to set the "congestion indication" bit in each packet.

The network is analyzed as a feedback control system. The solution adopted
is composed of several elements:

-- Routers set the congestion indication bit, in response to their average
queue length over a cycle + the current cycle (i.e. busy, idle, current busy
cycle).

-- The user updates the window size after receiving a number of ACKs equal
to the sum of the previous and current window size, but only the most recent
bits corresponding to the current window size are examined.

-- If at least half of these bits are set, then the window size is reduced
multiplicatively (0.875), otherwise it is increased additively (1.0).

The authors show that the scheme converges to an efficient operating point,
is fair, stable at high load, and fares well during transient changes to the
network.

Throughout the paper, the authors define several useful measures of
congestion such as Power and Efficiency. The paper does a good job of
documenting the reasoning for rejecting alternatives such as using
hysteresis for triggering the setting of a bit or adjusting the window size
after each ACK. The arguments are convincing and are further strengthened
by simulation results. The only apparent drawback with this approach is the
requirement for a change to the transport protocol implementations.
Received on Tue Sep 26 2006 - 08:07:19 EDT