Review - Congestion Avoidance and Control

Review of the paper "Congestion Avoidance and Control"
by Ivan Hernandez

The paper starts describing the dramatic performance degradation
because of the explosive growth of computer networks. The main cause
of this problem is the poor implementation of transport
protocol.

Thus, to solve this problem, the authors looked for pieces of a TCP
implementation that violate the conservation principle -- If the
network accomplish this principle, the network should be robust in
face of congestion. They stated three ways of violate the principle
either (1) the connection does not get to equilibrium (this
equilibrium is required by the principle), or (2) a sender injects a
new packet before and old packet has exited, or (3)the equilibrium can
not be reached because of resource limits along the path.

In order to solve the issues stated above, the authors add a new
per-connection state variable on the sender: the congestion window
(cwnd). To solve (1), to reach the required equilibrium, they
developed an algorithm that increases the packets a sender sends in a
moderate pace.At this point, the authors fix the round trip time
estimator, a key element in the TCP implementation in order to support
heavy load, they achieve this task calculating one if its parameters
instead of using a fixed value; this will lead to trustful timeout
information. Now, to solve (3) they developed a congestion avoidance
algorithm that decreases the cwnd to half its current value, and in
case of (ack x rtt) increases in one the cwnd. In both algorithms, the
sender sends the minimum of the receiver's advertised window and cwnd.

The paper compares its new TCP implementation with a previous
one. The new one shows a great improvement in the utilization of the
network.

This is a good paper. It shows a good theoretical and practical
balance, and their graphs are very illustrative. It is interesting the
authors' knowledge from areas such as physics and control theory to
support their ideas. As in previous papers reviewed, one of the goals
of this one is to get rid of the oscillation of the important values,
in this case used in the congestion solution.

In the future work section of the paper, the authors highlight that
the algorithms in the transport endpoints insure the network capacity
is not exceeded, they cannot insure fair sharing of capacity. So they
see gateways as convergence points with enough information to control
sharing and fair allocation, so they view the gateway congestion
detection algorithm as the next big step. Furthermore, they state that
the gateways should reduce congestion even if no endnode is modified
to do congestion avoidance. I think is against the e2e argument,
moreover, the gateways would have to implement L4 tasks.
Received on Mon Sep 25 2006 - 16:36:57 EDT