Dynamic Programming as a Scheduling Tool in Multiprogrammed Computing Systems

Abstract

A potentially parallel iterative algorithm for the solution of the unconstrained N-stage decision problem of Dynamic Programming is developed. This new solution method, known as Variable Metric Dynamic Programming, is based on the use of variable metric minimisation techniques to develop quadratic approximations to the optimal cost function for each stage. The algorithm is applied to various test problems, and a comparison with an existing similar algorithm proves favourable. The Variable Metric Dynamic Programming solution method is used in the implementation of an adaptive highlevel scheduling mechanism on a multiprogrammed computer in a university environment. This demonstrates a practical application of the new algorithm. More importantly, the application of Variable Metric Dynamic Programming to a scheduling problem illustrates how Mathematical Programming may be used in complex computer scheduling problems to provide in a natural way the required dynamic feedback mechanisms.