Providing high-availability, elasticity for an in-memory database system with RAMCloud

Abstract

Stanford's RAMCloud is a large-scale storage system that keeps all data in DRAM and provides high availability as well as a great degree of elasticity. These properties make it desirable for being used as the persistence for an in-memory database system. In this paper, we experimentally demonstrate the high availability and elasticity RAMCloud can provide when it is being used as a storage system for a relational in-memory database system: a) We utilize RAMCloud's fast-crash-recovery mechanism and measure its impact on database query processing performance. b) We evaluate the elasticity by executing a sinus-shaped, a plateau, and an exponential database workload. Based on our experiments, we show that an in-memory database running on top of RAMCloud can within seconds adapt to changing workloads and recover data from a crashed node - both without an interruption of the ongoing query processing.