Performance analysis of the reserve capacity policy for dynamic VM allocation in a SaaS environment

Abstract

We consider a periodic-review provision scheme with constant inspection intervals for allocating dynamically virtual machines (VMs) in a Software-as-a Service (SaaS) environment. At each interval, we determine how many virtual machines (VMs) to provisioned or de-provision using a simple heuristic referred to as the reserve capacity policy, since it maintains a fixed reserve capacity of VMs. We analyze the performance of the reserve capacity policy within the context of a periodic-review provision scheme using a Markov Chain embedded at the inspection intervals. We assume a single stream of jobs with each job requiring a single VM. Jobs arrive in a Poisson fashion and the execution time of a job in a VM is exponentially distributed. We calculate the probability distribution of the number of customers in the system, the number of in-service VMs, the utilization, and the queue-length distribution of the waiting customers. The embedded Markov Chain is solved numerically. For cases where the underlying transition matrix is very large, we have proposed approximations and showed that they have a root mean square error (RMSE) of less than 2%.