Hybrid Dissemination: Adding Determinism to Probabilistic Multicasting in Large-Scale P2P Systems

Epidemic protocols have demonstrated remarkable scalability and robustness in disseminating information on internet-scale, dynamic P2P systems. However, popular instances of such protocols suffer from a number of significant drawbacks, such as increased message overhead in push-based systems, or low dissemination speed in pull-based ones. In this paper we study push-based epidemic dissemination algorithms, in terms of hit ratio, communication overhead, dissemination speed, and resilience to failures and node churn. We devise a hybrid push-based dissemination algorithm, combining probabilistic with deterministic properties, which limits message overhead to an order of magnitude lower than that of the purely probabilistic dissemination model, while retaining strong probabilistic guarantees for complete dissemination of messages. Our extensive experimentation shows that our proposed algorithm outperforms that model both in static and dynamic network scenarios, as well as in the face of large-scale catastrophic failures. Moreover, the proposed algorithm distributes the dissemination load uniformly on all participating nodes.