Analysis of the Increase and Decrease Algorithm for Congestion Avoidance in Computer Network

Continuous growing in the Internet and mixing between various technologies with different speed (like twisted pair and fiber optic), result in congestion problem that degrades the network performance as response time and throughput. After the Cliff point where the load reaches the network capacity the throughput falls rapidly, the packets is being dropped and the time response increases drastically.

While tradition congestion control schemes recovers the network from such state, this papers describes a congestion avoidance control mechanism to guide the network to operate within the optimal range close to Knee point where the load resulted in a small increase in throughput with a reasonable increase in response time. Such a mechanism depends on the ability of the users to deal with dynamic window size or rate during the communication. This mechanism depends on the increase/decrease algorithm which in turn represents dynamic resource management. Network resources sense its load, determine if they are overloaded or under loaded and feedback simply a binary signal to their users to adjust their number of packets transmitted over the network. Window size in the transport layer protocol or rate of transmission of a user often are used to adjust the system load. The binary feedback signal (1=overload, 0=underload) will be represented by a bit in the packet header, and interpreted by the user as 0;increase load and 1;decrease load

The paper concentrates on the analysis of various classes of increase/ decrease algorithm where the user's reaction to the received feedback can be additive with a constant (positive or negative) or multiplicative by a constant (>1 or 0< .. >1). Also, It uses many performance metrics and the assumption of many users share the same one resource with synchronous feedback and control loop. These metrics includes;
· Efficient usage of the resource that can be defined by the closeness of the total load to Knee point
· Fair allocation of the resource between the users
· Fast convergence; time to reach the goal state and size of oscillation. The binary feedback prevents the convergence to single steady state and leads to an oscillation

It has been found that additive increase and multiplicative decrease algorithm satisfies the goal.

However, I don't have an idea if they thought to use a feedback signal with two bits instead of one it is very simple idea
First bit as mentioned before to represent the required reaction by the user either increase or decrease its load, and the second bit to inform the user to consider the defined constant or its half value
Ex: feedback 10 means decrease load by B/2
feedback 11 means decrease load by B
etc.
I think it performs better and the feedback still simple