In the client/server model for distributed on-line requests processing, the concept of a server class that consists of multiple identical servers is often provided for the fast response time, high fault-tolerance, or continuous availability. As far as we are aware of, there is no concrete formulation that provides the optimal number of servers of a server class in distributed client/ server systems. The number of servers has been usually chosen only by heuristic methods or statistics after monitoring system behavior for some periods of time. In this paper, we propose a method based on queuing theory that determines the optimal number of servers in a server class for a general distributed client/server model that encompasses the state-of-the-art on-line transaction processing environment described in the literature. (C) 2002 Elsevier Science B.V. All rights reserved.