Improved interlock correction when solving layered queueing networks using decomposition

Layered Queueing networks are a common method for solving performance models of modern distributed computer systems that use blocking remote procedure calls. Several analytic methods exist to solve these networks, many of which use the method of decomposition to break the model up into smaller, more easily solved submodels. Analytic solutions that break up a model must take into consideration interlocking, which is a phenomena that arises when a single customer in one submodel is represented by more than one customer in another. Failing to correct for interlocking can result in large errors in the final solution. This paper revisits interlocking, as implemented in the analytic Layered Queueing Network Solver. The interlock calculation it uses often distributes the waiting a customer experiences incorrectly among intermediate tasks. Further, certain models with external contention can yield unfeasible utilizations at interlocked servers. This paper introduces a new interlock calculation which is more accurate, and does not produce unfeasible utilizations. The new approach is compared against the old approach (and against solutions with no interlock correction) and is shown to produce better results in all cases.