Institute for Systems Research

In this dissertation, we consider the behavior of generalized TCP congestion avoidance when subject to randomized congestion feedback, such as RED and ECN. The window distribution of individual flows under a variable packet loss/marking probability is established and studied to demonstrate the desirability of specifying a less drastic reduction in the window size in response to ECN-based congestion feedback.

A fixed-point based analysis is also presented to derive the mean TCP window sizes (and throughputs) and the mean queue occupancy when multiple such generalized TCP flows interact with a single bottleneck queue performing randomized congestion feedback.

Recommendations on the use of memory (use of weighted averages of the past queue occupancy) and on the use of "drop biasing" (minimum separation between consecutive drops) are provided to reduce the variability of the queue occupancy.

Finally, the interaction of TCP congestion avoidance with randomized feedback is related to a framework for global optimization of network costs. Such a relation is used to provide the theory behind the shape of the marking (dropping) functions used in a randomized feedback buffer.

We first present an analyticaltechnique to obtain the 'mean' queue occupancy and the 'mean' of the individual TCP windows. Armed with this estimate of the means, wethen derive the window distribution of each individual TCPconnection. Extensive simulation experiments indicate that, under a wide variety of operating conditions, our analytical method is quite accurate in predicting the 'mean' as well asthe distributions. The derivation of the individual distributions is based upon a numerical analysis presented which considers the case of a single TCP flow subject to variable state-dependent packet loss.