Review - TCP Congestion Control with a Misbehaving Receiver

Congestion control mechanisms are used to throttle network connection rates.
This provides a mechanism for two machines to calibrate the communication
link, so that it can reach speeds that reflect network conditions and
performance of the end nodes. Savage et al. illustrate that methods inherent
to the TCP protocol can be used by a misbehaving receiver to increase the
connection rate.

This can be achieved through several acknowledgement attacks that serve to
quickly increase the sender's maximum segment size, or reduce the perceived
round trip time. Scenarios include ACK division, which is effective during
slow start, ACK duplication, which induces fast retransmit, and optimistic
ACKing, which can cause the TCP protocol to perceive a very short RTT. These
scenarios are possible because separate interests are not assumed by default
in the TCP protocol. It is a sharing protocol, where both participating
nodes are assumed to have a mutual interest in the wellbeing of the network
as a whole.

While the consequences of these flaws are presented clearly, the proposed
solutions do not seem to be well evaluated. Firstly, the implications of
inserting new elements into the TCP header are not trivial. TCP is used to
drive the Internet and requiring everyone to modify their implementation is
unrealistic. It would be much more feasible to create a new transport
protocol that could run in parallel. Furthermore, the performance
implications of these suggestions are not considered. Inserting more random
number calculations into the TCP header would likely reduce performance.

This paper is interesting and thought provoking. Clearly, many of the
suggestions for improvements outlined in this paper have not been
implemented in the six years since publication. This leads me to wonder what
improvements, if any, have been implemented to correct these issues?
Received on Sun Oct 30 2005 - 10:24:41 EST