Abstract
A content delivery network (CDN) typically consists of geographically-distributed data centers (DCs), which are deployed within proximity to end-users who request Internet content. Content copies are stored at the DCs and are delivered to end-users in a localized manner to improve service availability and end-to-end latency. On one hand, CDNs have improved QoS experienced by end-users. On the other hand, the rapid increase in Internet traffic volume has caused the global DC industry's energy usage to skyrocket. Therefore, our focus in this thesis is to realize energy awareness in CDN management while assuring end-to-end QoS. First, we surveyed the literature on energy-aware DC and CDN management schemes. We highlighted the significance of dynamically provisioning server and network resources in DCs in order to reduce DC energy usage. We also recognized that in order to achieve optimal CDN energy saving, energy optimization should be performed both within each DC and among multiple DCs in a CDN. Second, we proposed a theoretical framework that minimizes server power consumption in cross-domain CDNs. The term "server" refers to any co-locating entity that can handle user requests, e.g., server clusters or DCs. Our strategy was to put a subset of servers to sleep mode during off-peak hours to save energy. In order to avoid deteriorated QoS caused by less live server resources, we enforced constraints on utilization of servers and network links respectively to avoid them from being overloaded. Third, we designed an energy-aware CDN management system. The strategy was not only to put a subset of servers within each DC to sleep, but also to put entire DCs to sleep during off-peak hours through load unbalancing among DCs. We showed how the proposed system can be integrated within a typical modern CDN architecture. We also developed a heuristic algorithm that allows CDN operators to quickly make decisions on server and DC sleeping, as well as energy-aware request resolution. QoS was assured through constraints on server response time and end-to-end delay. Fourth, we built an optimization model that minimizes the overall energy consumption of CDN DCs, including their servers and cooling systems. We derived a lower bound to its optimal objective. Through comparing with the lower bound, we showed that our earlier developed heuristic algorithm's energy-saving gain was guaranteed to be near-optimal. We also quantitatively studied the trade-off between CDN energy saving and QoS performance in terms of end-to-end delay and server response time.