Optimal Pricing and Capacity Planning of a New Economy Cloud Computing Service Class

Resource under-utilization in cloud computing systems is widespread due to workload fluctuations and drives up the cost of cloud computing service. Offering service using slack resources in an opportunistic way improves the utilization of resources and the economics of cloud service providers. But Opportunistic service class comes with virtually no service level objectives (SLO) and thus is of limited use. In a recent study, a new Economy class was introduced to provide long-term SLOs using reclaimed cloud computing resources. Analysis based on the workload collected on six production cloud computing clusters at Google demonstrated the potential of the Economy class. This paper presents an analytic study on the optimal pricing and capacity planning of this new Economy class. We show that depending on the terms of the service level agreements and the characteristics of the cloud computing workloads, a cloud service provider may either choose a penalty averse or penalty preference strategy when allocating reclaimed computing resources to the Economy class cloud computing service. We also derive conditions under which the new Economy class will be profitable.

I. Introduction

Infrastructure-as-a-Service (IaaS) cloud provides users with affordable and elastic computing service. Essential to this affordability and elasticity are virtualization technology and statistical multiplexing. Virtualization technology enables the cloud service operators to provision virtual machines (VMs) instead of physical servers to host different applications. Each VM is allocated a certain amount of resources and multiple VMs can be placed on the same physical server. Statistical multiplexing exploits the reduction of the variability of aggregated workload fluctuations and allows the cloud operator to provision physical resources that are less than users' total requests for resources. Workload statistics collected from six Google production cloud computing clusters from December 2012 to November 2013 help make this point [1].