High Performance Transactions in Deuteronomy
Conference on Innovative Data Systems Research (CIDR 2015) |
The Deuteronomy architecture provides a clean separation of transaction functionality (performed in a transaction component, or TC) from data management functionality (performed in a data component, or DC). In prior work we implemented both a TC and DC that achieved modest performance. We recently built a high performance DC (the Bw-tree key value store) that achieves very high performance on modern hardware and is currently shipping as an indexing and storage layer in a number of Microsoft systems. This new DC executes operations more than 100× faster than the TC we previously implemented. This paper describes how we achieved two orders of magnitude speedup in TC performance and shows that a full Deuteronomy stack can achieve very high performance overall. Importantly, the resulting full stack is a system that caches data residing on secondary storage while exhibiting performance on par with main memory systems. Our new prototype TC combined with the previously re-architected DC scales to effectively use 48 hardware threads on our 4 socket NUMA machine and commits more than 1.5 million transactions per second (6 million total operations per second) for a variety of workloads.