Geographic Trough Filling for Internet Datacenters

TL;DR

This paper introduces energy-efficient trough filling strategies for datacenters that leverage delay-tolerant jobs to optimize energy use and delay performance, using dynamic speed scaling and traffic shifting.

Contribution

It presents two novel joint speed scaling and traffic shifting schemes that operate with minimal statistical information, improving energy efficiency and delay management in datacenters.

Findings

Significant energy savings demonstrated with proposed schemes

Effective delay performance achieved through load shifting

Schemes validated on real and artificial datacenter traces

Abstract

To reduce datacenter energy consumption and cost, current practice has considered demand-proportional resource provisioning schemes, where servers are turned on/off according to the load of requests. Most existing work considers instantaneous (Internet) requests only, which are explicitly or implicitly assumed to be delay-sensitive. On the other hand, in datacenters, there exist a vast amount of delay-tolerant jobs, such as background/maintainance jobs. In this paper, we explicitly differentiate delay-sensitive jobs and delay tolerant jobs. We focus on the problem of using delay-tolerant jobs to fill the extra capacity of datacenters, referred to as trough/valley filling. Giving a higher priority to delay-sensitive jobs, our schemes complement to most existing demand-proportional resource provisioning schemes. Our goal is to design intelligent trough filling mechanisms that are energy efficient and also achieve good delay performance. Specifically, we propose two joint dynamic speed scaling and traffic shifting schemes, one subgradient-based and the other queue-based. Our schemes assume little statistical information of the system, which is usually difficult to obtain in practice. In both schemes, energy cost saving comes from dynamic speed scaling, statistical multiplexing, electricity price diversity, and service efficiency diversity. In addition, good delay performance is achieved in the queue-based scheme via load shifting and capacity allocation based on queue conditions. Practical issues that may arise in datacenter networks are considered, including capacity and bandwidth constraint, service agility constraint, and load shifting cost. We use both artificial and real datacenter traces to evaluate the proposed schemes.