Modeling Peer-to-Peer Connections over a Smartphone Network

Abstract

Smartphones offer a natural platform for building decentralized systems for the common good. A very important problem in such systems is understanding the limitations of building a peer-to-peer (P2P) overlay network, given that today's networking infrastructure is designed with centralized services in mind. We performed measurements over smartphones over several years and collected large amounts of data about, among other things, P2P connection success. Here, we train models of P2P connection success using machine learning based on several features that are observable by the devices. We argue that connection success is a non-trivial function of many such features. Besides this, the predictive models are also rather dynamic and a good model can perform rather badly if it is based on data that is more than a year old. The degree distribution of the P2P network based on this model has an interesting structure. We can identify two modes that roughly correspond to "very closed", and "average"nodes, and a rather long tail that contains relatively open nodes. Our model allows us to perform realistic simulations of very large overlay networks, when combined with device measurement traces. This enables us to have a more informed design of decentralized applications.