(no subject)

Summary of the paper

This paper proposes a scheme for congestion avoidance for networks
using connectionless protocols at the network layer.

The main contribution of the paper work is the design of the new
scheme, using a binary feedback from the network, with just one bit
in the network layer header to indicate congestion. Authors modeled
the network as a feedback control system and identified the various
components of the scheme in the context of such a model.

At first, paper gives the distinction of the basic principle: flow
control and congestion control. One is a "selfish" control function,
and the other is a "social" problem. In the view of control theory,
the difference is which point the system wants to reach. The proposed
congestion avoidance policy that drives the operation of the network
toward the knee of the delay curve, rather than recovering from
operation near the cliff.

The design of the scheme is the binary feedback for congestion
avoidance. A router that is congested sets a congestion indication
bit in the network layer header of a data packet that is flowing in
the forward direction. When the data packet reaches the destination,
the congestion indication bit is copied into the transport layer
header of the acknowledgment packet. This acknowledgment packet is
then transmitted from the destination to the source. In this way, the
scheme uses the minimal amount of the feedback. This will not place
more traffic on the network which may probably cause congestion
collapse.

In fact, the feedback system has two sets of policies for controlling
the traffic placed on the network. One is "Router Policy", the other
one is "User Policy". For addressing these problems, paper firstly
builds the model and solution methodology. Authors define a function
called "Power" at each router, and use this function in the order to
choose the operating point of the network so that it can be at the
knee of the delay curve. To use the power at each resource to finally
determine the network operation point, authors use a function called
"Efficiency". The maximally efficient operating point for the
resource is its knee. Furthermore, to consider the fairness across
all the users of the network, authors provide the formula to measure
the fairness of any other nonequal allocation, and use throughputs to
measure allocations. The goal is to use the network efficiently,
while achieving fairness.

Then comes to the problems that how to detect the congestion, and
what's the algorithm to generate the feedback signal. The answer in
the paper is that monitoring the average queue length at the router
to detect congestion and using both the hysteresis and the single-
threshold policies. For feedback filter, to ensure that the system
operates at the correct point, author does not use the instantaneous
queue sizes at the router, but instead use the average queue size.
Moreover, it needs an adaptive averaging algorithm, which in effect,
determines the busy/idle periods adaptively at the router.

For policies for decision making, authors analyze it through 4
aspects: decision frequency, use of received information, signal
filtering, and increase/decrease algorithms. Users could determine
the correct window size to use according to these policies. Finally,
paper is followed by testing at various cases and conclusion.

Points in favour or against

The paper is generally well written, with fine and clear
presentation. The view points in this paper are different from the
paper "Congestion Avoidance and Control". There are several
differences: it's in the network layer, it mainly focuses on the
performance and congestion avoidance in the routers, and so on. But
it has relation with the former paper. The congestion avoidance in
the network layer will influence the decision on transport layer, and
support the algorithm in transport layer which is described in the
former paper. So I think, the congestion avoidance needs the
cooperation of several layers and different lays have different
objects. Furthermore, I think the unique view of this paper is that
authors model the network as a feedback control system. In this way,
the problem could be solved by some of control theories which could
give it a strong mathematic support. Another viewpoint I want to
indicate is that this paper considers the fairness problem. I think
it's really a big problem in the computer networks which most
literatures did not solve. Although in this paper, author did not
discuss it in much detail, it's a good viewpoint to support the
conclusion of this research work. However, in another hand, it's the
weak point of this paper, author should discuss the optimization
criteria in more detail and give more analysis in the latter parts of
the paper.
Received on Mon Sep 25 2006 - 17:38:39 EDT