Discrete event fluid modeling of background TCP traffic

Abstract

TCP is the most widely used transport layer protocol used in the Internet today. A TCP session adapts the demands it places on the network to observations of bandwidth availability on the network. Because TCP is adaptive, any model of its behavior that aspires to be accurate must be influenced by other network traffic. This point is especially important in the context of using simulation to evaluate some new network algorithm of interest (e.g., reliable multicast) in an environment where the background traffic affects---and is affected by---its behavior. We need to generate background traffic efficiently in a way that captures the salient features of TCP, while the reference and background traffic representations interact with each other. This article describes a fluid model of TCP and a switching model that has flows represented by fluids interacting with packet-oriented flows. We describe conditions under which a fluid model produces exactly the same behavior as a packet-oriented model, and we quantify the performance advantages of the approach both analytically and empirically. We observe that very significant speedups may be attained while keeping high accuracy.