Beyond Simple Request Processing with RAMCloud

RAMCloud is a scale-out data center key-value store designed to aggregate the DRAM of thousands of machines into petabytes of storage with low 5 μs access times. RAMCloud was an early system in the space of low-latency RDMA-based storage systems. Today, it stands as one of the most complete examples of a scalable low-latency store; it has features like distributed recovery, low fragmentation through a specialized parallel garbage collector, migration, secondary indexes, and transactions. This article examines RAMCloud’s key networking and dispatch decisions, and it explains how they differ from other systems. For example, unlike many systems, RAMCloud does not use one-sided RDMA operations, and it does not partition data across cores to eliminate dispatch and locking overheads. We explain how major RAMCloud functionality still exploits modern hardware while retaining a loose coupling both between a server’s internal modules and between servers and clients by explaining our recent work on Rocksteady, a live data migration system for RAMCloud. Rocksteady uses scatter/gather DMA and multicore parallelism without changes to RAMCloud’s RPC system and with minimal disruption to normal request processing.