ENERGY EFFICIENCY Traffi c-Aware Resource Provisioning for Distributed Clouds

loud-computing-based traffi c has been rapidly growing in recent years. Cisco forecasted that annual global datacenter IP traffi c will reach 7.7 zettabytes by the end of 2017, with its cloud IP traffi c reaching 5.3 zettabytes.1 Correspondingly, the service providers, including Google, Microsoft, Facebook, and AT&T, are building and expanding their datacenters nationwide and worldwide. Such geographically distributed datacenters are often referred to as Internet datacenters (IDCs) and we use cloud as a general term that refers to an IDC’s collection of hardware, software, and services. An IDC typically consumes many megawatts of power, which imposes a signifi cant electricity cost to its operator. For example, Google’s datacenters consume nearly 300 million watts annually.2 To reduce energy costs, researchers have proposed load-aware server provisioning schemes, in which the number of active servers is controlled dynamically based on the load.3–6 When the load is low, extra servers can be scheduled in sleeping mode. In this paradigm, obtaining traffi c volume information is a challenging issue. As the “Related Work in Resource Provisioning” sidebar describes, many researchers have worked on traffi c-aware cloud resource provisioning. However, considerable room for achieving a fi ne-grained traffi c-awareness remains. In the load-aware server provisioning schemes,4–7 researchers typically consider traffi c dynamics in a large time scale only, such as tens of minutes, during which traffi c demand (that is, input to the server provisioning algorithms) is usually assumed static given the current time interval for server provisioning. However, as we observed from a real