Fast Paxos Made Easy

Abstract

Distributed consensus is one of the most important building blocks for distributed systems. Fast Paxos is one of the latest variants of the Paxos algorithm for distributed consensus. Fast Paxos allows an acceptor to cast a vote for a value of its choice unilaterally in a fast round, thereby eliminating a communication step for reaching consensus. As a tradeoff, the coordinator must build a quorum that is bigger than the simple majority used in Classic Paxos. This article presents the theory, implementation, and a comprehensive performance evaluation of the Fast Paxos algorithm. The theory is described in an easier-to-understand way compared with the original article by Lamport. In particular, an easy-to-implement value selection rule for the coordinator is derived. In the implementation of Fast Paxos for state-machine replication, a number of additional mechanisms are developed to cope with practical scenarios. Furthermore, the experiments reveal that Fast Paxos is most appropriate for use in a single-client configuration. The presence of two or more concurrent clients even in a local area network would incur frequent collisions, which would reduce the system throughput and increase the mean response time as experienced by clients. Due to frequent collisions, Fast Paxos actually performs worse than Classic Paxos in the presence of moderate to large number of concurrent clients.