Simple and Effective Dynamic Provisioning for Power-Proportional Data Centers

TL;DR

This paper investigates how future workload information impacts dynamic server provisioning in data centers, developing online algorithms with proven competitive ratios and demonstrating their effectiveness through simulations with real-world data.

Contribution

It introduces a novel divide-and-conquer approach to the offline problem, designs simple online algorithms with competitive guarantees, and shows future workload beyond a critical window does not improve performance.

Findings

Optimal offline provisioning structure characterized

Online algorithms with competitive ratios up to approximately 1.58 achieved

Future workload information beyond a critical window does not enhance performance

Abstract

Energy consumption represents a significant cost in data center operation. A large fraction of the energy, however, is used to power idle servers when the workload is low. Dynamic provisioning techniques aim at saving this portion of the energy, by turning off unnecessary servers. In this paper, we explore how much performance gain can knowing future workload information brings to dynamic provisioning. In particular, we study the dynamic provisioning problem under the cost model that a running server consumes a fixed amount energy per unit time, and develop online solutions with and without future workload information available. We first reveal an elegant structure of the off-line dynamic provisioning problem, which allows us to characterize and achieve the optimal solution in a {}"divide-and-conquer" manner. We then exploit this insight to design three online algorithms with competitive ratios $2-\alpha$, $(e-\alpha)/(e-1)\approx1.58-\alpha/(e-1)$ and $e/(e-1+\alpha)$, respectively, where $0\leq\alpha\leq1$ is the fraction of a critical window in which future workload information is available. A fundamental observation is that \emph{future workload information beyond the critical window will not} \emph{improve dynamic provisioning performance}. Our algorithms are decentralized and are simple to implement. We demonstrate their effectiveness in simulations using real-world traces. We also compare their performance with state-of-the-art solutions.