Numerical Methods for Solution of Queuing-Network Problems with Applications to Models of Multiprogrammed Computer Systems.

摘要

The major difficulties in determination of the steady-state properties of a Markovian queuing network by numerical solution of a set of linear balance equations are the choice of vector representation of the states, the generation and storage of the states, and generation, storage and solution of the balance equations. Lexicographic sequencing of the vector representations are shown in this thesis to lead to efficiencies in the storage and solution of the balance equations and to provide a key to efficient generation and storage of the states. Convergence properties of three iterative solution methods are examined for cyclic models, such as those which can result from a central-server model. An analysis of possible bias in software monitors on computer systems is analyzed in terms of a central-server model of such systems. Techniques for examining tape-mounting policies and core-allocation policies are also suggested. (Author)