Managing life-cycle capacity of cloud computing: Integrating data-driven optimization and inventory theory for capacity investment and retirement

Abstract

The burgeoning cloud computing market prompts cloud service providers (CSPs) to offer efficient and potent services, contingent upon the judicious allocation of service resources. The pitfalls of overprovisioning and underprovisioning—wasting cloud resources and diminishing service performance, respectively—result in avoidable economic disparities. This paper delves into the cloud service management conundrum, seeking to uncover the most efficient dynamic investment and retirement strategies throughout the entire life cycle of a cloud service product, encompassing the growth and the decline stages. We formulate three progressive scenarios based on different levels of demand knowledge: deterministic demand, stochastic demand, and a distribution-free demand data. We demonstrate the preserved optimality of time-dependent $(s_t,S_t)$ policies with stochastic demand, and introduce a high-quality Adaptive-Approximation-Algorithm-Based policy that assures a performance guarantee of 3. We also construct a data-driven framework by employing predictive prescriptive methods to execute online investment and retirement strategies with distribution-free demand data. Empirical evidence from Huawei Cloud corroborates that these predictive prescriptive approaches markedly enhance the efficacy of cloud service management, achieving cost reductions and optimized distribution rates.