Review. Di Niu

Review: A Binary Feedback Scheme for Congestion Avoidance in Computer
Networks

Reviewer: Di Niu

This paper proposes a binary feedback scheme for congestion avoidance
in computer networks. It involves a series of algorithms that help
operate the network at an optimal point, called the knee of the delay
curve. The basic rationale behind this scheme is to retrieve from
preceding packet transmissions the feedback about the congestion
informations along the paths, and have users adjust their window
sizes accordingly to avoid congestion. While this idea of using
feedback was actually shared by many other works at that time [1],
[2], [13], [14], [18] (references in this paper), the paper
distinguishes itself by its meticulous tuned engineering details and
parameters. Highlights among these details include a binary feedback
rather than packet-based feedback, the consideration of fairness, the
distributed decision making, the careful filtering of feedback
information at the congestion generating points, and the refined
window size adjustment at end users who are going to utilize this
feedback.

A bit-based feedback is used in order to avoid exacerbating
congestion, as a packet-based scheme would have done. Great care has
been taken to filter congestion information before it is fedback to
the users. The idea of the filter is to only keep load information
at a router that lasts long enough for the user action to be
meaningful. Thus, the average queue length is considered.

According to the paper, the decision making at a user has to decide
how often to update the window size, which and how feedback
information should be used, and how to adjust the window size
accordingly. It turns out this paper came up with a window size
adjustment algorithm (Algorithm B) quite similar to the other paper
for review ("Congestion Avoidance and Control"). Additive increase
and multiplicative decrease in window size are adopted. As a step
further with respect to the other paper, "Congestion Avoidance and
Control", fairness is considered in choosing various window
adjustment algorithms. In addition, in evaluating the discrepancy in
the resource allocation among different users, a good fairness
criterion is used, which could be adopted in many other fairness
problems such as load balancing in P2P networks.

Despite of the delicacy in this feedback congestion avoidance scheme,
it is still questionable whether feedback from the intermediate
router is useful and practical in a large network. Although the
parameters are meticulously tuned in this paper, yet real networks
can have an extremely heterogeneous traffic pattern where the schemes
in this paper may not work at all. As a simple example, since the
feedback information is always outdated information when it is used,
it is hard to evaluate how old it is, and thus very hard to obtain
useful information from it through filtering. In other words, the
schemes in this paper may only be suitable to specific network
environment.
Received on Mon Sep 25 2006 - 22:57:15 EDT