vRaft: Accelerating the Distributed Consensus Under Virtualized Environments

In recent years, Raft has been gradually widely used in many distributed systems (e.g., Etcd, TiKV, PolarFS, etc.) to ensure the distributed consensus because it is effective and easy to implement. However, because the performance of the virtual nodes in cloud environments is usually heterogeneous and fluctuant due to the “noisy neighbor” problem and the cost efficiency, the strong leader mechanism makes the Raft protocol encounter a serious performance challenge. Specifically, when the performance of the leader node is low, the whole system performance will descend accordingly since both the write and the read requests serving will be blocked by the slow leader processing. Aiming to solve this problem, we proposed a modified version of Raft specially optimized for virtualized environments, i.e., vRaft. It breaks Raft’s strong leader restriction and can fully utilize the temporarily fast followers to accelerate both the write and the read requests processing in a virtualized cloud environment, without affecting the linearizability guarantee of Raft. The experiments based on the virtual nodes in Tencent Cloud indicate that vRaft improves the throughput by up to 64.2%, reduces average latency by 38.1%, and shortens the tail latency by 88.5% in a typical read/write-balanced workload compared with Raft.