In this paper we describe our simulation models for evaluating end-to-end performance of HTTP transactions. We first analyze several gigabytes of collected traffic traces from a production Frame Relay network. Using these traces we extract web traffic and analyze the server web pages accessed by actual users. We analyze over 25000 web pages and develop a web client/server interaction model based upon our analysis of many server contents. We make specific contributions by analyzing the popularity of web servers and the number of bytes transferred from them during a busy hour. We also compute the distribution of the number of embedded items within a web document. We then use these models to drive a network simulation and show the interactions between the TCP/IP flow control and retransmission mechanism on source parameters. One of our important contributions has been to show that the Hurst parameter is robust with regard to TCP/IP flow and error control. Using the simulation studies we show that the distribution of end-to-end application message delay has a heavy-tail distribution and we discuss how these distributions arise in the network context.