High performance state-machine replication

Scalable State-Machine Replication

“…Other works have tried to make linearizable systems scalable (CORBETT et al, 2013;GLENDENNING et al, 2011;MARANDI;PEDONE, 2011). In (GLEN-DENNING et al, 2011), the authors propose a scalable key-value store based on DHTs, …”

“…Previous optimistic replicated designs (e.g., (JIMÉNEZ-PERIS et al, 2002;KEMME et al, 1999;KOTLA et al, 2007;MARANDI;PEDONE, 2011)) require replicas to recover from commands executed in the wrong order, when the optimistic and the conservative orders do not match. Fast-SSMR does not have this constraint: in case of order mismatches, a correct state is copied from the conservative state machine to the optimistic one.…”

“…The baseline Fast-SSMR is inspired by previous optimistic designs to reduce the latency of replication (e.g., (JIMÉNEZ-PERIS et al, 2002;KEMME et al, 1999;KOTLA et al, 2007;MARANDI;PEDONE, 2011) Optimistic replicated designs require the ability to recover from commands executed in the wrong order. Previous approaches have tackled this issue by relying on the recoverability of the environment in which the service is deployed (e.g., transaction…”

S-Paxos: Offloading the Leader for High Throughput State Machine Replication

“…Recently, it has been considered in the context of state machine replication [26], [27]. In [26], Marandi et al show that state partitioning leads to improvements in both throughput and response time for a replicated B-tree service. However, as they argue, perfect state partition is often not possible.…”

“…One can gain performance even with imperfect state partitioning, as it lifts the requirement that every request needs to be received by every replica. In [26] the authors use Ring-Paxos as an ordering protocol. However, as they point out in [27], when the number of partitions increases (which is required for good throughput), Ring-Paxos becomes the bottleneck due to its leader-centric nature.…”