RubbleDB: CPU-Efficient Replication with NVMe-oF.

Due to the need to perform expensive background compaction operations, the CPU is often a performance bottleneck of persistent key-value stores. In the case of replicated storage systems, which contain multiple identical copies of the data, we make the observation that CPU can be traded off for spare network bandwidth. Compactions can be executed only once, on one of the nodes, and the already-compacted data can be shipped to the other nodes' disks, saving them significant CPU time. In order to further drive down total CPU consumption, the file replication protocol can leverage NVMe-oF, a networked storage protocol that can offload the network and storage datapaths entirely to the NIC, requiring zero involvement from the target node's CPU. However, since NVMe-oF is a one-sided protocol, if used naively, it can easily cause data corruption or data loss at the target nodes. We design RubbleDB, the first key-value store that takes advantage of NVMe-oF for efficient replication. RubbleDB introduces several novel design mechanisms that address the challenges of using NVMe-oF for replicated data, including pre-allocation of static files, a novel file metadata mapping mechanism, and a new method that enforces the order of applying version edits across replicas. These ideas can be applied to other settings beyond key-value stores, such as distributed file and backup systems. We implement RubbleDB on top of RocksDB and show it provides consistent CPU savings and increases throughput by up to 1.9x and reduces tail latency by up to 93.4% for write-heavy workloads, compared to replicated key-value stores, such as ZippyDB, which conduct compactions on all replica nodes.