Energy-efficient disk caching for content delivery

Content delivery networks (CDNs) operate hundreds of thou sands of servers that cache content and deliver them to users with high performance. Each CDN server has multiple spinning disks that are used for caching content. These disks account for 40-55% of the total server energy usage of a CDN. Reducing the energy consumption of a CDN by shutting down some of the disks during off-peak hours is the main focus of our work. The primary challenge with this approach is that shutting down disks decreases the size of the content cache, potentially lowering the cache hit rates, and resulting in a degradation in user-perceived performance. Our main contribution is developing and evaluating algorithms for cache sizing, disk shutdown, content placement and eviction that allow disks to be shut down without significantly impacting cache hit rates and user-perceived performance. We empirically evaluate the energy-performance tradeoff for our algorithms using extensive request traces from the world's largest CDN. We show that it is feasible to obtain a 30% disk energy savings with a 6.5% decrease in the normalized server hit rate and a mere 3% reduction in the normalized cluster hit rate. This work establishes disk shutdown as a key mechanism for energy savings in CDNs, paving the way for its future rollout on production networks.