Internet content distribution

This thesis studies Internet content distribution. In the first part of the thesis we consider client redirection mechanisms. We develop an architecture for locating copies of cached objects. This architecture is a small extension to the Domain Name System and can be deployed incrementally. We present an architecture of an Internet-wide replicated directory service and show how the current Domain Name System can be implemented with this architecture. The key features of this architecture are that it allows us to store rapidly changing information, can be deployed incrementally, and requires no changes to existing software. Our extensive performance evaluation of this architecture provides us with insight on how long the information can be cached. We also evaluate the performance of the redirection mechanisms used by modern content distribution networks. We find that the overhead of opening new connections to new servers can severly limit the user-perceived performance. In the second part of the thesis we consider object replication in content distribution. We develop a combinatorial optimization model for optimally replicating objects in a content distribution network. Our results show that best performance is obtained when replication is coordinated over the whole network. Using the same model we also develop cooperation strategies for peer-to-peer networks. We also consider the problem of optimal content replication in peer-to-peer communities. We formulate this problem as an integer programming problem and develop several adaptive algorithms to replicate objects on-the-fly. Our results indicate that our algorithms combined with least-frequently-used replacement policy provide near-optimal performance. We also consider the distribution of layered encoded video using a stochastic knapsack model. We develop several heuristics to determine which layers of which videos should be cached in order to maximize the accrued revenue.