Content Management in Planet-Scale Video CDNs

Global video Content Distribution Networks (CDNs) serve a significant fraction of the entire Internet traffic through a global network of cache servers. Effective handling of the massive traffic at the edge is vital for the feasibility of these CDNs, which can otherwise incur significant monetary costs and resource overloads in the Internet. We study this problem from several angles. First, we design cache management algorithms for individual servers of these CDNs: an LRU-based baseline solution to address their unique requirements; a flexible ingress-efficient algorithm; a greedy offline cache aware of future requests to estimate the maximum possible efficiency; and an optimal offline cache (for limited scales). In addition, we study cross-server content management. We analyze how in the absence (impracticality) of cooperative caching, the knowledge of requests in other serving locations can lead to better caching decisions overall and reducing up to hundreds of Gbps of costly traffic. We call this practice cache coordination and design the proper mechanism for realizing it. Based on actual workload data from a large-scale, global video CDN, we first conduct a detailed analysis of the spatial correlation of video popularities---workload similarities---worldwide. We then analyze the effectiveness and feasibility of cache coordination and its scalability: from within a city to across countries. Furthermore, we study the problem of provisioning the right CDN server cluster to deploy in each location, a central problem for today's continuously expanding CDNs. We optimize the proper server count, peering bandwidth, and server configuration as the disk drives and the necessary SSD and/or RAM cache layers to sustain the intensive I/O load. The optimization framework carefully captures the interaction of cache layers inside each server, the interplay between egress/disk capacity and network bandwidth, storage drives' read/write constraints and their costs. Finally, we study the generic problem of large-scale message distribution in networked systems with minimum end-to-end latency. We develop a rich suite of algorithms for different message delivery requirements and application scenarios, with up to 60% reduction in delivery latency and support for several times larger scales compared to existing solutions.Ph.D