How Geo-Distributed Data Centers Do Demand Response: A Game-Theoretic Approach

We study the demand response (DR) of geo-distributed data centers (DCs) using smart grid's pricing signals set by local electric utilities. The geo-distributed DCs are suitable candidates for the DR programs due to their huge energy consumption and flexibility to distribute their energy demand across time and location, whereas the price signal is well-known for DR programs to reduce the peak-to-average load ratio. There are two dependencies that make the pricing design difficult: 1) dependency among utilities; and 2) dependency between DCs and their local utilities. Our proposed pricing scheme is constructed based on a two-stage Stackelberg game in which each utility sets a real-time price to maximize its own profit in Stage I and based on these prices, the DCs' service provider minimizes its cost via workload shifting and dynamic server allocation in Stage II. For the first dependency, we show that there exists a unique Nash equilibrium. For the second dependency, we propose an iterative and distributed algorithm that can converge to this equilibrium, where the “right prices” are set for the “right demands.” We also verify our proposal by trace-based simulations, and results show that our pricing scheme significantly outperforms other baseline schemes in terms of flattening the power demand over time and space.

I. Introduction

Data centers (DCs) are well-known as large-scale consumers of electricity (e.g., DCs consumed 1.5% of the worldwide electricity supply in 2011 and this fraction is expected to grow to 8% by 2020 [1]). A recent study shows that many DC operators paid more than $10M [2] on their annual electricity bills, which continues to rise with the flourishing of cloud-computing services. Therefore, it is necessary for DC operators to both cut costs and increase performances. Recent works have shown that DC operators can save more than 5%–45% [3] operation cost by leveraging time and location diversities of electricity market prices to optimize geo-distributed DCs. However, most of the existing research is based on one important assumption: the electricity price applying to DCs does not change with demand. This assumption may not be true since an individual DC with enormous energy consumption (e.g., Facebook’s DC in Crook County, Oregon can contributed up to 50% of the total load of its distribution grid [4]) will impact to the supply demand balance of its local utility, which in turn can alter the utility’s price as shown in recent studies [5]–[7]. Furthermore, the power grid can be negatively affected due to this assumption. For example, blackouts might happen due to overloads in these areas where the DCs operator shifts all of its energy demand to a local utility with a low price and a high enough background load.