How to dynamically optimize the deployment of an application server cluster according to the load condition to balance the power and performance is an important issue that must be urgently solved. This paper proposes an online real-time optimization strategy for power-performance tradeoff for application server clusters, which aims to minimize the weighted value of the power and request discarding rate of a cluster. The optimization content involves the on/off state and CPU frequency of each server. The strategy involves two methods: small-scale cluster optimization(SSCOpt) and large-scale cluster optimization(LSCOpt). The SSCOpt defines a large number of variables to describe cluster optimization as a linear mixed integer programming problem, and then uses a software package to solve the problem. By analyzing the traits of power and load models, the LSCOpt defines a small number of variables to describe cluster optimization as a nonlinear mixed integer programming problem, and then proposes an solution algorithm based on flower pollination algorithm and variable fusion. The experimental results show that, the SSCOpt can get the global optimal deployment rapidly when used in small-scale clusters, and the LSCOpt can find a good near-optimal deployment rapidly even when applied to large-scale clusters.