Efficient Consensus and Synchronization for Distributed Systems

Abstract

Recent interest in decentralized applications calls for distributed systems that replicate their states across a large number of servers communicating over wide-area networks. We propose near-optimal solutions to two fundamental problems in the design and implementation of such systems: consensus and synchronization. First, we propose a universal decomposition of distributed consensus protocols that enables near-optimal throughput and liveness on fluctuating networks. Our key technique is to minimize the amount of communication necessary for participants to safely reach consensus. We design a state-of-the-art information dispersal protocol to achieve that. Second, we propose the first family of efficient rateless error-correcting codes for reconciling set differences. Our codes enable pairs of servers to synchronize system states with near-optimal communication and computation costs. We theoretically analyze these solutions, and implement end-to-end systems to demonstrate strong real-world benefits.