Cloud overbooking through stochastic admission controller

Abstract

Cloud providers apply overbooking to increase utilization of data center resources, which is associated with the risk of overloading cloud resources. In this paper, we study an admission control technique that permits cloud overbooking with bounded probability of resource over-utilization. The objective is to achieve a specified quality-of-service related to the probability of resource over-utilization in an uncertain loading condition, while maintaining high resource utilizations. Our method relies on estimating the probability of over-utilization based on approximating the probability distribution of the total resource demand on hosts as Beta. We perform a qualitative study to investigate the efficiency of using our method on Google Compute Cluster where we disclose some empirical observations on how well the resource utilization can be estimated as Beta Distribution. We also report results on the performance of the stochastic admission controller by estimating the mean and the standard deviation of the usage data.