Don’t mix pages with di erent lifetimes in one stream

We present a demonstration about optimizing two database storage engines by leveraging multi-streamed SSDs (MS-SSD in short). By storing data pages with similar lifetime together in the same physical ash blocks, MS-SSD can e ectively reduce the overhead of garbage collection, improving the write performance and prolonging the lifespan. Thus, in order to bene t from MS-SSD, it is very crucial for database storage engines to precisely classify logical data pages according to their update intervals and to e ectively map those logical data streams to physical streams in MS-SSD. Given that numerous new interfaces between host and ash memory SSD for better performance are emerging, this demonstration will provide a model case of physical database tuning on ash memory SSDs. We have successfully multi-streamed two database engines, MySQL/InnoDB and ForestDB, by identifying several logical data streamswith distinct update intervals in each engine and by taking the stream-per-object policy, instead of the naïve streamperle one. By running both vanilla and multi-streamed versions of two storage engines on real MS-SSD, we showcase that multi-streamed versions consistently outperform vanilla ones. In addition, we propose a set of guidelines on how to group logical streams with di erent update intervals into smaller number of physical streams with minimal performance degradation.