Failure Recovery from Persistent Memory in Paxos-Based State Machine Replication

Abstract

Paxos is one of the most popular protocols for state machine replication (a technique used for making services highly available). We are the first to propose a Paxos-based state machine replication framework which is aimed at persistent (non-volatile) memory, pmem in short-a new class of memory offering direct byte-addressable access to memory (e.g., Optane ™ DC Persistent Memory). In the paper, we describe two variants of the framework, called mPaxosSM and mPaxos, which support efficient recovery of processes after crash with the use of pmem. In the latter variant, a part of Paxos's state, and in the former also the entire state machine's state that should survive crashes, are stored in the persistent memory. This allows to achieve low failure recovery time. We used a key-value map to compare our frameworks equipped with different memory backends (pmem, DRAM, and emulated pmem), with the classical Paxos that recovers state from snapshots and logs stored in stable storage, and with Paxos equipped with EpochSS-a state-of-the-art protocol ensuring state recovery from peer replicas. Our results show the advantages of pmem and our approach.