Kernel Paxos

Abstract

State machine replication is a well-known technique to build fault-tolerant replicated systems. The technique guarantees that replicas of a service execute the same sequence of deterministic commands in the same total order. At the core of state machine replication is consensus, a distributed problem in which replicas agree on the next command to be executed. Among the various consensus algorithms proposed, Paxos stands out for its optimized resilience and communication. Much effort has been placed on implementing Paxos efficiently. Existing solutions make use of special network topologies, rely on specialized hardware, or exploit application semantics. Instead of proposing yet another variation of the original Paxos algorithm, this paper proposes a new strategy to increase performance of Paxos-based state machine replication. We introduce Kernel Paxos, an implementation of Paxos that significantly reduces communication overhead by avoiding system calls and TCP/IP stack. To reduce the number of context switches related to system calls, we provide Paxos as a kernel module. We present a detailed performance analysis of Kernel Paxos and compare it to a user-space equivalent implementation.