A Hierarchical Model for Fast Distributed Consensus in Dynamic Networks

Abstract: We present two new consensus algorithms for dynamic networks. The first, Fast Raft, is a variation on the Raft consensus algorithm that reduces the number of message rounds in typical operation. Fast Raft is ideal for fast-paced distributed systems where membership changes over time and where sites must reach consensus quickly. The second, C-Raft, is targeted for distributed systems where sites are grouped into clusters, with fast communication within clusters and slower communication between clusters. C-Raft uses Fast Raft as a building block and defines a hierarchical model of consensus to improve upon throughput in globally distributed systems. We prove the safety and liveness properties of each algorithm. Finally, we present an experimental evaluation of both algorithms in AWS.