We evaluate response times, in N-user collaborations, of the popular centralized (client-server) and replicated (peer-to-peer) architectures, and a hybrid architecture in which each replica serves a cluster of nearby clients. Our work consists of definitions of aspects of these architectures that have previously been unspecified but must be resolved for the analysis, a formal evaluation model, and a set of experiments. The experiments are used to define the parameters of and validate the formal analysis. In addition, they compare the performances, under the three architectures, of existing data-centric, logic-centric, and stateless shared components. We show that under realistic conditions, a small number of users, high intra-cluster network delays, and large output processing and transmission costs favor the replicated architecture, large input size favors the centralized architecture, high inter-cluster network delays favor the hybrid architecture, and high input processing and transmission costs, low think times, asymmetric processing powers, and logic-intensive applications favor both the centralized and hybrid architectures. We use our validated formal model to make useful predictions about the performance of the three kinds of architectures under realistic scenarios we could not create in lab experiments.