Dynamic power management in content distribution networks

The Internet was designed for communication between hosts but is increasingly used as a means to distribute content. Therefore, overlay networks are required for high-quality, highly available, and bandwidth-efficient content distribution. Such content distribution networks (CDNs) consist of cache servers deployed in data centers at the edge of the Internet close to the content consumers. A cache server replicates the most popular content from the origin server, which stores the original content. A CDN redirects a client request from the origin server to an available cache server close to the client. The growing consumer demand for quality-sensitive video streaming services on a broadening range of mobile terminals spurs the expansion of content distribution networks towards the consumers. Internet service providers (ISPs) have started to deploy cache servers in data centers in their own regional networks to offer multiscreen IPTV services with guaranteed quality to paying subscribers. A CDN owned by an ISP is called a telco CDN. The rapidly rising number of power-hungry cache servers fuels the unsustainable increase of data-center power consumption. Cache servers consume much power in part to keep their disks spinning. In this thesis, we aim to save energy in content distribution networks. We approach this problem in four steps: (1) a survey and analysis of power-reduction techniques for data storage systems, (2) a CDN energy simulator, (3) a near-optimal offline policy for combined server and disk power control, and (4) an online server and disk power-control policy. Because cache servers are typically equipped with many disks, a CDN can be considered a large-scale data storage system distributed across multiple data centers. Therefore, we survey all of the existing techniques that save energy in data storage systems such that we can identify the techniques applicable to CDNs. We classify the power-reduction techniques according to the disk-power factor and storage-stack layer addressed. For each technique, we reveal the fundamental trade-offs between power, capacity, performance, and dependability. Most of the power-reduction techniques are based on dynamic power management (DPM), which turns off underutilized system components. To clarify DPM at the level of the disk under a data-center workload, we analyze this technique based on an analytical model for the energy consumed by a disk during an idle period. To evaluate potential energy-saving techniques, we propose a CDN energy simulator driven by workload traces recorded by an operational telco CDN delivering IPTV services to mobile devices by means of HTTP adaptive streaming. Existing CDN simulators simulate only performance not power consumption. Our simulator models the power consumption of the cache server's disks separately from the server itself such that disks and servers can be turned on and off independently. Simulations show that CDNs are not energy proportional, which means their energy consumption is not proportional to their load. Therefore, the diurnal load fluctuations revealed by the traces lead to energy waste. Energy can be saved by applying DPM to the cache servers. To save energy in CDNs, we propose applying DPM at the level of the servers and their disks. We are the first to consider such combination of server and disk power management. So far only server power management was applied to CDNs. To determine an upper bound for the energy savings that can be realized by means of DPM, we present a near-optimal offline greedy heuristic power-control policy. This power-control policy targets maximum energy savings while maintaining high availability and bandwidth efficiency. We evaluate this power-control policy using our CDN energy simulator. Even for a minimally provisioned CDN, 28% of the energy can be saved. Finally, we present an online load-directed threshold-based policy for combined server and disk power control in CDNs. Similar to the offline power-control policy, the online policy reduces the CDN's energy dissipation while ensuring high availability and bandwidth efficiency. In addition, the proposed policy avoids an excessive wear of servers and disks. The evaluation of the online policy is based on the CDN energy simulator. Even when provisioning the minimum number of cache servers and disks required to cope with the daily peak load, the online policy reduces the power consumption by 24%. Such energy savings correspond to 85% of the energy-savings upper bound determined by the offline policy. Moreover, the energy savings realized by the online policy that combines server and disk power management are 30% larger than the energy savings achieved by the existing online policy limited to server power management.