Log-Based Data Replication With Consistency Guarantee For Distributed Persistent Memory Systems

Persistent Memory(PM)provides data persistence and achieves DRAM-like performance along with higher capacity,which offers the opportunity to improve the overall performance of distributed systems in datacenters.Since datacenter environments suffer from potential network failures or hardware and software failures in servers,distributed PM systems need data replication to add redundancy for reliability and availability.Meanwhile,data should be replicated across multiple nodes to divide the heavy workloads from clients.However,directly applying existing replication algorithms to distributed PM systems will cause double writes,small random writes and the read bottleneck,which weaken the lowlatency and high-throughput PM.In order to avoid double writes and small random writes with atomic data replication while providing local reads to eliminate the read bottleneck,this paper proposes a log-based data replication scheme for distributed PM systems,called RLog(Replicated LogStructured Persistent Memory).RLog concurrently supports appending logs in a lock-free way to avoid double writes with data atomicity in replication.By committing multiple updates in a batch,RLog reduces small random accesses to PM.For the fast indexing of persistent log structure,RLog proposes a highly concurrent index,which can support independent inter-key accesses.RLog designs a replication protocol with consistency guarantee.RLog commits the replication in only a single round trip,thus achieving lowlatency and high-throughput data replication for distributed PM systems.RLog keeps backup servers invalid before committing to defend clients from uncommited data,thus providing the linearizable local reads and non-blocking reads to improve the read performance of replicated systems.Experimental results exhibit that compared with the state-of-the-art replication scheme,Log Shipping,RLog on average achieves 37.2% higher throughput,17.5% lower median latency and 8.8% lower tail latency;For even more comprehensive evaluation,compared with the Non-Atomic Replication in Hermes,RLog achieves the near performance on throughput,median latency and tail latency.