A Markov decision process model for computer system load control

This paper combines a queueing network representation with a Markov decision process model to obtain optimal load control policies in a computer system with batch and interactive job classes. A precise queueing network model for the system is shown to lead to a computationally prohibitive state space, and consequently, an approximation based on treating the memory loop as a closed queueing network is developed. The approximate system is solved numerically under several parameter configurations using policy iteration. The performance of the optimally controlled system is compared to the uncontrolled system with a simulation model. The sensitivity of the performance to the system parameters is investigated.