LevelDB-Raw: Eliminating file system overhead for optimizing performance of LevelDB engine

摘要

Key-value store, a kind of NoSQL database, is data storage system for storing and retrieving key-value pairs using unique keys. In this paper, we present LevelDB-Raw, which improves LevelDB, one of the most well-known key-value store engines. LevelDB-Raw opens a raw block device file directly, bypassing the existing file system layer. Instead, it implements a simple, light-weight user-level file system to eliminate file system overhead such as redundant journaling and metadata management. We demonstrate the performance improvement of LevelDB-Raw using ForestDB-Benchmark and YCSB workloads. The results indicate that LevelDB-Raw increases the overall performance by 1.16x-3.45x in HDD and 1.05x-2.26x in SSD for updating database.