This thesis describes the design and implementation of state machine replication (SMR) that achieves near-perfect load balancing and availability, near-optimal request processing latency (especially in the wide area), and performance robustness when confronted with failures and slow replicas.
Traditionally, practical replicated state machines have used leader-based implementations of consensus algorithms, because it has been believed that they provide the best performance---highest throughput and lowest latency. At the same time, however, a leader-based approach has many drawbacks: the failure of the leader halts the entire replicated state machine temporarily, the speed of the entire set is determined by the speed of the leader, and, in geo-replicated scenarios, the distance to the leader causes remote clients to experience high latency.
This work shows that leaderless approaches can not only solve these problems and provide the flexibility of a completely decentralized system, but they can also achieve substantially higher performance than leader-based protocols. We introduce a new variant of the Paxos protocol that we call Egalitarian Paxos. In Egalitarian Paxos all replicas perform the same functions simultaneously to ensure better load balancing and availability, lower commit latency and higher performance robustness when compared to previous Paxos variants. We show---both theoretically and empirically---that Egalitarian Paxos has the aforementioned benefits when updating the state of a replicated state machine. We then apply the same leaderless design principle to improve the SMR read performance: quorum read leases generalize previously proposed time lease-based approaches to allow arbitrary sets of replicas to perform highly consistent local reads for parts of the replicated state.
David Andersen (Chair)
Michael Kaminsky (Intel Labs)
Miguel Castro (Microsoft Research)
deb [atsymbol] cs ~replace-with-a-dot~ cmu ~replace-with-a-dot~ edu