Back To The Future: On Predicting User Uptime

Abstract

Correlation in user connectivity patterns is generally considered a problem for system designers, since it results in peaks of demand and also in the scarcity of resources for peer-to-peer applications. The other side of the coin is that these connectivity patterns are often predictable and that, to some extent, they can be dealt with proactively. In this work, we build predictors aiming to determine the probability that any given user will be online at any given time in the future. We evaluate the quality of these predictors on various large traces from instant messaging and file sharing applications. We also illustrate how availability prediction can be applied to enhance the behavior of peer-to-peer applications: we show through simulation how data availability is substantially increased in a distributed hash table simply by adjusting data placement policies according to peer availability prediction and without requiring any additional storage from any peer.