Cost- and Energy-Aware Load Distribution Across Data Centers Kien Le † , Ricardo Bianchini † , Margaret Martonosi ‡ , and Thu D. Nguyen † † Rutgers University ‡ Princeton University 1 Introduction Today, many large organizations operate multiple data centers. The reasons for this include natural business dis- tribution, the need for high availability and disaster toler- ance, the sheer size of their computational infrastructure, and/or the desire to provide uniform access times to the infrastructure from widely distributed client sites. Re- gardless of the reason, these organizations consume sig- nificant amounts of energy and this energy consumption has both a financial and environmental cost. Interestingly, the geographical distribution of the data centers often exposes many opportunities for optimizing energy consumption and costs by intelligently distribut- ing the computational workload. We are interested in three such opportunities. First, we seek to exploit data centers that pay different and perhaps variable electric- ity prices. In fact, many power utilities now allow con- sumers to choose hourly pricing, e.g. [1]. Second, we seek to exploit data centers that are located in different time zones, which adds an extra component to price vari- ability. For example, one data center may be under peak- demand prices while others are under off-peak-demand prices. Third, we seek to exploit data centers located near sites that produce renewable (hereafter called “green”) electricity to reduce “brown” energy consumption that is mostly produced by carbon-intensive means, such as coal-fired power plants. To make our investigation of these degrees of free- dom more concrete, in this paper we consider multi-data- center Internet services, such as Google or iTunes. These services place their data centers behind a set of front-end devices. The front-ends are responsible for inspecting each client request and forwarding it to one of the data centers that can serve it, according to a request distribu- tion policy. Despite their wide-area distribution of re- quests, services must strive not to violate their service- level agreements (SLAs). This paper proposes and evaluates a framework for optimization-based request distribution. The framework enables services to manage their energy consumption and costs, while respecting their SLAs. It also allows services to take full advantage of the degrees of freedom mentioned above. Based on the framework, we propose two request distribution policies. For comparison, we also propose a greedy heuristic designed with the same goals and constraints as the other policies. Operationally, an optimization-based policy defines the fraction of the clients’ requests that should be di- rected to each data center. The front-ends periodically (e.g., once per hour) solve the optimization problem de- fined by the policy. After fractions are computed, the front-ends abide by them until they are recomputed. The heuristic policy operates quite differently. During each hour, it first exploits the data centers with the best power efficiency, and then starts exploiting the data centers with the cheapest electricity. Our evaluation uses a day-long trace from a commer- cial service. Our results show that the optimization- based policies can accrue substantial cost reductions by intelligently leveraging time zones and hourly electricity prices. The results also show that we can exploit green energy to achieve significant reductions in brown energy consumption for small increases in cost. Related work. The vast majority of the previous work on data center energy management has focused on a single data center. We are not aware of any previous work that addresses load distribution across data centers with respect to their energy consumption or energy costs. Moreover, we are not aware of other works on leverag- ing time zones, variable electricity prices, or green en- ergy sources. The exception here is [10], which leverages electricity price diversity to shut down entire data centers when their electricity costs are relatively high. Finally, we know of no previous work on optimization-based re- quest distribution in Internet services, besides our own [8]. However, our previous work did not address energy issues, time zones, or heuristics at all. 2 Request Distribution Policies We assume that a front-end is chosen to first handle a client request via round-robin DNS or some other high- level policy. The front-ends execute one of our policies and forward each request to a data center that can serve it. Typically, a request can only be served by 2 or 3 mirror data centers; further replicating content would increase the state-consistency traffic without a meaningful benefit 1