Presending is an active service which extends caching mechanism from temporal locality to spatial locality. Two modes of extracting user behavior patterns are proposed to predict future requests from clients for efficient presending. URL-based mode exploits the Markov-chain features of request series, and can be used for hierarchical presending. Session-based mode captures more semantics, and the authors' work emphasizes the clustering algorithm, feasible document weight definition, and attribute-vector-distance computation representing order of accesses. Their performance is evaluated using appropriate metrics such as request hit rate, session hit rate, presending efficiency and presending cost. Numerous experiments are carried out to compare the two modes. These methods are used for web presending, while they are helpful to web server design and ISP (internet service provider) service planning.