Abstract: Methods of pure performance testing or single analytical modeling, such as queueing network model, etc, have their limitation on the accuracy of performance indexes measurement, the validity of performance forecasting, and the controlling of testing iteration due to the complexity of Web systems. A Web performance modeling framework supporting mixed performance modeling is proposed. It uses different performance modeling methods for different kinds of performance indexes to derive closed form functions and their hypothesis of measurement. The regression analysis and testing are used on the training data to estimate the parameters of the closed form functions. To demonstrate the feasibility and validity of this framework, a real-world Web community system (igroot.com) is studied under the framework. For the indexes of system response time and scalability, a mixed modeling method is proposed by combining queueing network reduction and extended universal scalability model US-γ. Compared with other practical system performance testing methods, such as universal scalability model US, the model accuracy of performance forecasting is greatly improved and the cost of software and hardware used in the process is greatly reduced. The error rate of estimated response time is within 4 percent, the error rate of estimated throughout saturation point is within 1 percent, and the error rate of estimated infimum of buckle point is within 5 percent. Correlating the scalability model and threads data of the Web server, an HTTP processing bottleneck at the architecture level is identified.