Paper Review : Analysis of the Increase and Decrease Algorithms for Congestion Avoidance in Computer Networks
Reviewer : Seh Leng Lim
With congestion avoidance mechanisms, a network can operate in the optimal region of low delay and high throughput (that is at the knee of its throughput/load curve), thereby avoiding being congested. The paper focuses on first-order increase/decrease algorithms for congestion avoidance, using metrics such as efficiency, fairness, convergence time, and size of oscillation to compare them. By comparison, the authors are able to identify the sort of increase/decrease congestion algorithm that will help the network settle to an efficient and fair state regardless of its starting state.
The main contribution of the paper is its proposal that for optimal convergence to efficiency and feasibility, the increase policy should be additive and the decrease policy should be multiplicative. It does so by gently walking the reader through the relevant first order equations using the criterion of efficiency, fairness, distributedness and convergence.
The key main ideas expounded are:
(a) The binary feedback scheme whereby to keep the network operating in the optimal region of low delay and high throughput, the network monitors its load level and uses a bit in the network packet header to ask the users to increase or decrease the load as appropriate.
(b) An additive increase with multiplicative decrease window sizing policy will help to produce optimal convergence to fairness
I think that the paper has a significant contribution (rating of 4) to the study of congestion avoidance mechanisms in the computer networks. The paper does have an interesting approach to its comparison of the collision avoidance mechanisms with the use of algebraic mathematics. Logically and theoretically, the propositions in the paper are convincing. However, there is no experimental data to backup the claims except for the announcement that this will be implemented in the Digital Networking Architecture and OSI Transport Class 4 Networks. Nevertheless, it is still a very interesting and stimulating paper to read.
The authors have identified a few to be explored areas such as the impact of delayed feedback on control (since the authors have assumed that the feedback is synchronous for all users), the marginal utility of increased bits of feedback and the benefits of guessing the current number of users.
Researchers and builders who build TCP/IP stacks, routers or even internet applications will have a better feel of the limitations in implementing network congestion avoidance mechanisms from this paper.