A Binary Feedback Scheme for Congestion Avoidance in Computer Networks

Motivation: This paper proposes a scheme for congestion avoidance in
networks using a connectionless protocol at the network layer. The focus
is to use as small an additional field and as little bandwidth for
feedback as possible.

Key Points:

1) Congestion avoidance (operate the network at a highly desirable
point) and congestion control (detect the fact that network has reached
the cliff and then reduce the load to recover to an uncongested state)
has been distinguished.

2) The router that is congested sets the congestion avoidance bit in the
network level header of the data packet when the average queue length on
packet arrival is >=1. This bit is copied into the transport layer
header of the ACK packet and is then used by the source to modify its
window size.

3) Important to use average queue length instead of instantaneous to
provide a consistent state of the router and may also lead to unfairness.

4) If Wp is the window size before update and Wc window size after
update, ( Wp + Wc ) packets need to be acknowledged before the next
update is performed. A waiting period has been introduced as altering
window size after every ack causes considerable oscillation as well as
it takes time for the change to be able to reflect in the feedback.

5) Use only the relevant information ( last Wc bits ) while updating
window size to avoid any additional book-keeping about history of
congestion in the network.

6) Modification to window size is additive increase and multiplicative
decrease to ensure fairness. However, trial and error is used for coming
to the constant values which has not been fully explained.

7) Analysis shows that the approach is distributed, reasonably fair,
adapts to network changes, converges to efficient operating point and is
reasonably simple to implement with low operation overheads. The scheme
operates the network at a stable point during overload and with
arbitrary initial values for window sizes.

8) A disadvantage is that all routers need to be modified to adhere to
this policy. Instead, the congestion bit information could be already
deduced from timeouts at the source.

9) Its unclear in the paper how the regeneration point would be
calculated effectively given the fact that the network traffic widely
fluctuates and is characterized by bursts.

10) The scheme depends on sending feedback information coded in packets
at the time of congestion which itself adds to the problem of congestion
and also is unreliable in terms of packet reachability.
Received on Tue Sep 26 2006 - 04:48:44 EDT