A Stochastic Control Model for Leasing Computational Resources

Abstract

We present a model for determining the optimal resource leasing policy for a dynamic grid service. The model assumes that the demand for the service as well as the actual execution times are unknown, but can be estimated. We cast the problem in a Dynamic Programming framework and we are able to show that the model can make good resource leasing decisions in the face of such uncertainties. In particular, we use the model to decide how many resources should be leased for the service and for how long. The results show that use of the model reduces the cost of leasing computational resources and significantly reduces the variance of the cost.