Optimizing key-value stores for hybrid storage environments
Prashanth Menon · TSpace (University of Toronto) · 2015
Modern write-intensive key-value stores have emerged as the prevailing data storage system for many big applications. However, these systems often sacrifice their read performance to cope with high data ingestion rates. Solid-state drives (SSD) can lend their help, but the volume of data and the peculiar characteristics of SSDs make their exclusive use uneconomical. Hence, hybrid storage environments with both SSDs and hard-disk drives (HDD) present interesting research opportunities for optimization. This thesis investigates how to design and optimize key-value stores for hybrid storage environments. We first modify and extend an existing key-value store to leverage the SSD as a cache. We next formulate an analytical cost model to predict the performance of a generic log-structured hybrid system. Finally, we use insights from our model to design and implement a new hybrid-optimized key-value store, LogStore. We demonstrate that LogStore is up to 7x faster than LevelDB, a state-of-the-art key-value store.