As distributed systems are deployed across machines free to opt in or out at any time, host availability is a first-order concern. Systems typically deal with availability in a reactive way, making little distinction between machines with different availability behaviors. In short, current systems view fluctuating availability as a problem. This project views availability as an opportunity, by predicting the availability of many individual hosts in a larger collective. By using a competitive tournament of predictors, the project is able to predict the availability of many nodes even in completely unmanaged networks. The research effort applies these predictors in three domains. In the first, epidemic propagation, the project augments current viral infection models to consider availability in addition to topology. The second application is delay-tolerant networking. By using availability prediction as a practical contact oracle, one can improve end-to-end latency and routing efficiency. As a third application, the project applies availability prediction to DHT-based storage and archival networks. By biasing placement of some replicas towards highly available nodes, churn-induced document transfer is reduced and overall system availability improved. By ugmenting the maintenance protocols with availability predictions, network overhead is further reduced.
