Despite the growing popularity of user modeling servers, little attention has been paid to optimizing and evaluating the performance of these servers. We argue that implementation issues and their influence on server performance should become the central focus of the user modeling community, since there is a sharply increasing real-life load on user modeling servers, This paper focuses on a specific implementation-level aspect of user modeling servers – the choice of push or pull approaches to evidence propagation. We present a new push-based implementation of our user modeling server CUMULATE and compare its performance with the performance of the original pull-based CUMULATE server.