支持日志乱序提交的分布式一致性协议

Distributed consistency protocol supporting log submission out-of-order

为了解决Raft算法中日志提交的严格串行化设计,提出一种Raft协议变体:并行提交Raft (PC-Raft). PCRaft在日志提交阶段运用流水线,重新设计日志缓冲区,实现日志的乱序提交.传输方式使用RDMA网络,在提高日志传输的并发性的同时加快传输速度.在日志执行阶段,采用批处理,将多条指令打包发送给状态机逐条执行.针对日志并行提交情况下会出现的幽灵复现问题,采用LSN与任期号结合的方式保证日志的安全性.针对日志乱序提交会出现的日志空洞问题,重新设计领导者选举算法,在选举中加入临时领导者,保证选举出的领导者能最快恢复日志.测试结果证明PC-Raft对比Raft在吞吐量方面有着明显的性能提升,同时延迟更低,并且在日志指令依赖频繁的情况下,吞吐量比现有基于Raft的变体更高,延迟也更低.

A Raft protocol variant, parallel commit Raft(PC-Raft), was proposed, in order to solve the strict serialization design of log submission in Raft algorithm. In PC-Raft, pipeline was used in the log submission stage and the log buffer was redesigned to realize out-of-order submission of logs. The RDMA network was used as the transmission method to improve the concurrency of log transmission and speed up the transmission. In the log execution stage, batch processing was used to package multiple instructions and send them to the state machine for execution one by one. For the ghost recurrence problem that occurs, the combination of LSN and term number was used to ensure the security of the log. In order to solve the problem of log holes caused by disordered log submission,the leader election algorithm was redesigned, and temporary leaders were added to the election to ensure that the elected leaders can recover the logs as quickly as possible. Test results show that PC-Raft has a significant performance improvement in throughput compared with Raft, and the delay is lower. In the case of frequent log instruction dependency, the throughput is higher and the latency is lower compared with the existing Raft based variant.