A Quantitative Analysis of Redundancy Schemes for Peer-to-Peer Storage Systems

Fully decentralized peer-to-peer (P2P) storage systems lack the reliability guarantees that centralized systems can give. They need to rely on the system's statistical properties, only. Nevertheless, such probabilistic guarantees can lead to highly reliable systems. Moreover, their statistical nature makes P2P storage systems an ideal supplement to centralized storage systems, because they fail in entirely different circumstances than centralized systems. In this paper, we investigate the behavior of different replication and erasure code schemes as peers fail. We calculate the data loss probability and the repairing delay, which is caused by the peers' limited bandwidth. Using a Weibull model to describe peer behavior, we show that there are four different loss processes that affect the availability and durability of the data: initial loss, diurnal loss, early loss, and longterm loss. They need to be treated differently to obtain optimal results. Based on this insight we give general recommendations for the design of redundancy schemes in P2P storage systems.