Using ‘Drop–Biasing’ to Stabilize the Occupancy of Random-Drop Queues with TCP Traffic
Date: November 19 - November 23, 2000
The paper describes how the use of ‘drop-biasing’, a technique to control the distribution of the gap between consecutive packet losses in random drop queues (such as RED) can be used to reduce the variability of the queue occupancy with TCP traffic. Reducing the variance of the queue occupancy reduces delay jitter for buffered packets, as well as decreases the likelihood of buffer underflow. We find that modifying the packet drop probabilities to ensure a minimum separation between consecutive packet drops serves to decrease the variability in the queue occupancy. This is really achieved as a result of the increased negative correlation among the congestion windows of the constituent TCP flows. Such negative correlation explains why the use of simple drop-biasing strategies can reduce the queue variability without increasing the likelihood of bursts of packet losses. The results of our investigations have relevance for the design and deployment of RED-like algorithms for congestion control in the Internet.