content management systems. This paper describes how an (offline) experimental testbed may be used to address performance concerns. To begin, timing mea­ surements are collected from a server component incorporating the Policy Decision Point (PDP) under test, using representative policies and corresponding requests. Our experiments with two XACML PDP implementations show that measured request service times are typically clustered by request type; thus an algorithm for request cluster identification is presented. Cluster characterisation s are used as inputs to a PDP performance model for a given policy/request mix and an analytic (queueing) model is used to estimate the equilibrium server load for different mixes of request clusters. The analytic performance prediction model is validated and extended by discrete event simulation of a PDP subject to additional load. These predictive models enable network administrators to explore the capacity of the PDP for different overall loadings (requests per unit time) and profiles (relative frequencies) of requests.