Abstract

Abstract-Energy related costs are becoming one of the largest contributors to the overall cost of operating a data center, whereas the degree of data center utilization continues to be very low. Energy-aware dynamic provision of resources based on the consolidation of existing application instances can simultaneously address under-utilization of servers while highly reducing energy costs. Thus, energy costs cannot be treated separately from resource provision and allocation. However, current scheduling techniques based on market mechanisms do not specifically deal with such scenario. In this paper we model the problem of minimizing energy consumption of the allocation of resources to networked applications as a Stackelberg leadership game to find an upper bound of energy saving. The model is applied to a proportional-share mechanism where resource providers can maximize their profit by minimizing energy costs while users can select resources ensuring the minimum requirements are satisfied. We show that our mechanism can determine the optimal set of resources on and off, even in realistic conditions considering incomplete information, and heterogeneous applications. I. INTRODUCTION Energy consumption is a key concern in networked computing systems, including service overlays, content distribution systems and many other distributed systems. These systems require a collection of networked computing resources from one or multiple providers on data centers spread over the world. All data centers or cloud computing providers face the problems of high energy costs, a discouraging low server utilization, and changing resource demand of applications. For instance, Hoelzle et al. Two popular methods for effectively matching power consumption to workload requirements in a data center are via processor speed scaling and powering down nodes. The goal of this paper is to develop a theoretical framework to analyze the limits of energy saving, and strategies to determine the right set of computing nodes on and off to minimize energy consumption while keeping the right level of service for networked applications using these computational resources. We consider several characteristics of modern applications and resources -e.g. proportional share of resources, granularity of application processes, level of parallelism, and migration and consolidation of virtual machines. The focus and contributions of this paper are as follows: i) we derive a simple energetic model for computing elements based on SPEC benchmarks using real data; ii) we model