Efficient Content Distribution With Managed Swarms

Content distribution has become increasingly important as people have become more reliant on Internet services to provide large multimedia content. Efficiently distributing content is a complex and difficult problem: large content libraries are often distributed across many physical hosts, and each host has its own bandwidth and storage constraints. Peer-to-peer and peer-assisted download systems further complicate content distribution. By contributing their own bandwidth, end users can improve overall performance and reduce load on servers, but end users have their own motivations and incentives that are not necessarily aligned with those of content distributors. Consequently, existing content distributors either opt to serve content exclusively from hosts under their direct control, and thus neglect the large pool of resources that end users can offer, or they allow end users to contribute bandwidth at the expense of sacrificing complete control over available resources. This thesis introduces a new approach to content distribution that achieves high performance for distributing bulk content, based on managed swarms. Managed swarms efficiently allocate bandwidth from origin servers, in-network caches, and end users to achieve system-wide performance objectives. Managed swarming systems are characterized by the presence of a logically centralized coordinator that maintains a global view of the system and directs hosts toward an efficient use of bandwidth. The coordinator allocates bandwidth from each host based on empirical measurements of swarm behavior combined with a new model of swarm dynamics. The new model enables the coordinator to predict how swarms will respond to changes in bandwidth based on past measurements of their performance. In this thesis, we focus on the global objective of maximizing download bandwidth across end users in the system. To that end, we introduce two algorithms that the coordinator can use to compute efficient allocations of bandwidth for each host that result in high download speeds for clients. We have implemented a scalable coordinator that uses these algorithms to maximize system-wide aggregate bandwidth. The coordinator actively measures swarm dynamics and uses the data to calculate, for each host, a bandwidth allocation among the swarms competing for the host's bandwidth. Extensive simulations and a live deployment show that managed swarms significantly outperform centralized distribution services as well as completely decentralized peer-to-peer systems