Dynamic Bandwidth Allocation for Lifetime-Based TCP Classification

We study the separation of TCP flows into short lived and long lived ones, and the subsequent allocation of bandwidth to reflect the dynamically changing load due to short lived flow fluctuations. Three schemes are presented to assist in determining and allocating the bandwidth to short lived flows. Each one represents a different compromise in terms of state and execution complexity. However, all three share the common assumption that they can exploit knowledge of the statistical properties of what constitutes a “typical” short lived TCP flow. We argue that this assumption is totally within reason and hinges on the fact that short lived TCP flows are HTTP connections that have been extensively studied and their statistics are increasingly well understood. We finally present an evaluation of the three proposed schemes against DropTail and RED and determine that all three provide, to varying degrees, reduced response time to the short lived flows. Since short lived flows are predominantly instantiated by interactive applications, we argue that the reduced response time demonstrated by the proposed schemes is in good agreement with the most pressing performance concern of these applications, i.e., short delays.