RubbleDB - CPU-Efficient Replication with NVMe-oF

Explore a conference talk from USENIX ATC '23 that introduces RubbleDB, a novel key-value store designed for efficient replication using NVMe-oF. Learn how this innovative approach trades CPU usage for spare network bandwidth in replicated storage systems, executing compactions only once on one node and shipping the compacted data to other nodes' disks. Discover the challenges of using NVMe-oF for replicated data and the solutions implemented in RubbleDB, including pre-allocation of static files, a new file metadata mapping mechanism, and a method to enforce version edit order across replicas. Understand how these techniques can be applied beyond key-value stores to distributed file and backup systems. Examine the performance improvements achieved by RubbleDB, including increased throughput and reduced tail latency for write-heavy workloads compared to traditional replicated key-value stores like ZippyDB.