(no subject)

This paper lays out a number of congestion control techniques that I
believe are still in use today. The development and use of these
techniques is motivated by a historical failure in the ARPANET that
demonstrated weaknesses within BSD4.3's TCP implementation.
Depending upon which part of the paper you read, there are either
five main techniques - round-trip-time variance estimation,
exponential retransmit timer backoff, slow-start, more aggressive
receiver ack policy, and dynamic window sizing on congestion - plus
two more that are not described in the paper, or three - slow-start,
round-trip timing, and congestion avoidance. I consider this to be a
major organizational weakness of the paper in that the paper would be
much clearer if the introduction mirrored the organization of the
paper's body. Instead, the topics are introduced out of order, and
without regard for how they will later be grouped: variance
estimation and exponential backoff falling under round-trip timing,
and ack policy and dynamic window sizing falling under congestion
avoidance. In fairness, the second half of the introduction lays out
the structure of the paper... it is just that the connection between
this structure and the one laid out in the first half of the
introduction is never made clear.

Aside from the complaint above, however, the paper was incredibly
clear and well-written, involving a strong problem-based organization
structure on a high level and clear and readable sentences as well,
which is a rarity in academic papers.

Without reading their references, it is not clear how much of their
technique is novel and how much was based upon prior work; however, I
believe that much of their contribution consists of merging
complementary ideas from other researchers into to a complete
solution. What is impressive is that this paper lays out a number of
techniques that are still widely used, and describes a working
implementation of them working in parallel.

I am somewhat skeptical however, of their use of the term "congestion
avoidance". Their "avoidance" mechanism is triggered by packet loss,
which is a sign that congestion is already occurring, and hence it is
too late to avoid it. In a router-independent congestion control
scheme, no better indicator of congestion is available, so I am not
faulting their approach, only the terminology they use to describe it.

Just because it is so much fun to say, this is a clear example of end-
to-end design principles.
Received on Mon Sep 25 2006 - 21:00:48 EDT