To Reserve or Not to Reserve: Optimal Online Multi-Instance Acquisition in IaaS Clouds

TL;DR

This paper introduces two online algorithms for dynamically combining on-demand and reserved instances in IaaS clouds, achieving near-optimal cost savings without requiring future demand knowledge.

Contribution

It proposes the first practical online algorithms with proven optimal competitive ratios for multi-instance reservation in IaaS clouds, without needing demand distribution assumptions.

Findings

Algorithms achieve the best possible competitive ratios.

Significant cost savings demonstrated through simulations.

Algorithms extend to cases with reliable short-term predictions.

Abstract

Infrastructure-as-a-Service (IaaS) clouds offer diverse instance purchasing options. A user can either run instances on demand and pay only for what it uses, or it can prepay to reserve instances for a long period, during which a usage discount is entitled. An important problem facing a user is how these two instance options can be dynamically combined to serve time-varying demands at minimum cost. Existing strategies in the literature, however, require either exact knowledge or the distribution of demands in the long-term future, which significantly limits their use in practice. Unlike existing works, we propose two practical online algorithms, one deterministic and another randomized, that dynamically combine the two instance options online without any knowledge of the future. We show that the proposed deterministic (resp., randomized) algorithm incurs no more than 2-alpha (resp., e/(e-1+alpha)) times the minimum cost obtained by an optimal offline algorithm that knows the exact future a priori, where alpha is the entitled discount after reservation. Our online algorithms achieve the best possible competitive ratios in both the deterministic and randomized cases, and can be easily extended to cases when short-term predictions are reliable. Simulations driven by a large volume of real-world traces show that significant cost savings can be achieved with prevalent IaaS prices.