Optimal control of the degree of multiprogramming in a computer system

This paper deals with an optimal control problem in a multiprogrammed computer system. First, the quasi-birth-and-death process which models the system is investigated and an optimal static admission policy is derived that maximizes the throughput among policies with fixed maximum degree of multiprogramming. Numerical results show that this optimal static admission policy gives a ten percent higher throughput than the L=S and knee criteria. Next, the problem is formulated into a semi-Markov decision process and an optimal dynamic admission policy is obtained which controls the admission of jobs depending on the state of the system. It is shown numerically that this dynamic admission policy attains a five percent higher throughput than the optimal static admission policy in the system with a low-speed paging disk.