Self-Adaptive Cloud Capacity Planning

Abstract

The popularity of cloud service spurs the increasing demands of cloud resources to the cloud service providers. Along with the new business opportunities, the pay-as-you-go model drastically changes the usage pattern and brings technology challenges to effective capacity planning. In this paper, we propose a new method for cloud capacity planning with the goal of fully utilizing the physical resources, as we believe this is one of the emerging problems for cloud providers. To solve this problem, we present an integrated system with intelligent cloud capacity prediction. Considering the unique characteristics of the cloud service that virtual machines are provisioned and de-provisioned frequently to meet the business needs, we propose an asymmetric and heterogeneous measure for modeling the over-estimation, and under-estimation of the capacity. To accurately forecast the capacity, we first divide the change of cloud capacity demand into provisioning and de-provisioning components, and then estimate the individual components respectively. The future provisioning demand is predicted by an ensemble time-series prediction method, while the future de-provisioning is inferred based on the life span distribution and the number of active virtual machines. Our proposed solution is simple and computational efficient, which make it practical for development and deployment. Our solution also has the advantages for generating interpretable predictions. The experimental results on the IBM Smart Cloud Enterprise trace data demonstrate the effectiveness, accuracy and efficiency of our solution.