Consensus algorithms deployed in the crash fault tolerant setting chose a leader-based architecture in order to achieve the lowest latency possible. However, when deployed in the wide area they face two key robustness challenges. First, they lose liveness when the network is unreliable because they rely on timeouts to find a leader. Second, they cannot have a high replication factor because of the high load imposed on the leader-replica making it a bottleneck. This effectively limits the replication factor allowed, for a given level of throughput, thus lowering the fault tolerance threshold. In this paper, we propose RACS and SADL, a modular state machine replication algorithm that addresses these two robustness challenges. To achieve robustness under adversarial network conditions, we propose RACS, a novel crash fault-tolerant consensus algorithm. RACS consists of two modes of operations: synchronous and asynchronous, that always ensure liveness. RACS leverages the synchronous network to minimize the communication cost to O(n) and matches the lower bound of O(n2) at adversarial-case executions. To avoid the leader bottleneck and to allow higher replication factor, without sacrificing the throughput, we then propose SADL, a novel consensus-agnostic asynchronous dissemination layer. SADL separates client command dissemination from the critical path of consensus and distributes the overhead evenly among all the replicas. The combination of RACS and SADL (SADL-RACS) provides a robust and high-performing state machine replication system. We implement and evaluate RACS and SADL-RACS in a wide-area deployment running on Amazon EC2.
翻译:暂无翻译