In this paper we present a new model of Web traffic and its applications in the performance evaluation of Web servers. We consider typical behavior of a user's hypertext navigation within a Web server. We propose a traffic model at the session level, formulated as a stochastic marked point process, which describes when users arrive and how they browse the server. We provide results of statistical analyses and goodness-of-fit of various simple parametric distributions and of their mixtures. We developed a Web server benchmark: WAGON (Web trAffic GeneratOr and beNchmark), and we validated the traffic model by comparing various characteristics of the synthetic traffic generated by WAGON against measurements. We then report benchmark and analysis results on the Apache server, the currently most used Web server software. We analyze the impact of the traffic parameters on the HTTP request arrival process and the packet arrival process. We also show that the aggregate traffic is self-similar in most cases, and that, more importantly, the Hurst parameter is increasing in the traffic intensity. We further provide performance comparison results between HTTP1.0 and HTTP1.1 and show that HTTP1.1 could be much worse for users as well as for servers if some control parameters of the server and of browsers are incorrectly set. Indeed, when the server load is relatively high, the page response time under HTTP1.1 increases exponentially fast with the number of parallel persistent connections and with the timeout value used in Apache for persistent connections. We investigate the impact of user network conditions on the server performance. We also propose a queueing model to analyze the workload of persistent connections on the Apache server, and we establish optimal solution of the timeout parameter for the minimization of workload. Based on our analyses, we suggest the following practical guidelines. It is usually beneficial for both Web servers and Web clients to use HTTP1.1 instead of HTTP1.0. When HTTP1.1 is used, it should be used with pipeline. In terms of the management of persistent connections, it is useful for browsers to implement Early Close policy which combines the advantages of both HTTP1.0 and HTTP1.1. Browsers should in general, except for users with low bandwidth network connections (such as Modem), avoid establishing multiple parallel persistent connections from one browser window to the same Web server. On the server side, servers should set small timeout values for persistent connections if fixed timeout control mechanism is used (as in Apache) or if dynamic timeout control mechanism is used and the measured workload is high.