Computation of bounds and performance guarantees for client-server systems using a queueing network model

Multiclass queueing networks have been suggested by a number of researchers for the performance engineering of client-server systems. A number of implementation mechanisms, such as software servers, locks in transaction processing and critical sections, make the application of existing analytical methods for queueing networks difficult for the performance analysis of client-server software and systems. Moreover, real systems are unlikely to meet the underlying assumptions required by the exact solution methods that are available. We present a performance bounds technique which is apt for modelling client-server systems and is based on relaxed assumptions about the stochastic behaviour of the system. Models can be solved to give upper and lower bounds on throughputs for different classes of clients. Of particular interest is the lower bound on throughput that gives a performance guarantee that is useful in the conservative design of systems