REXIO: Indexing for Low Write Amplification by Reducing Extra I/Os in Key-Value Store Under Mixed Read/Write Workloads

In mixed read/write (r/w) workloads, Key-Value stores that utilize Log-Structured Merge trees face significant write amplification (WA) due to frequent writing, resulting in extensive data compaction. While current LSM-based KV stores effectively address WA in write-heavy workloads, they struggle with increased block cache invalidation and increased r/w I/O conflicts in mixed r/w workloads. This paper introduces REXIO, a novel indexing approach to reduce WA. It decouples RAM from Solid State Drives (SSDs) and stores addresses of KV pairs in the In-RAM hashing table, which reduces buffer invalidations and eliminates unnecessary data reorganization to lower extra I/Os, thereby reducing I/O conflicts. Additionally, REXIO stores keys and values in different blocks and employs In-blocking logging to optimize updating and deleting. The experimental results show that REXIO achieves a 68.3% reduction of WA and 3.4x throughput compared to state-of-the-art LSM-Tree approaches.

Recommended citation: Wang, Zizhao and Qu, Qiang and Han, Nan and Deng, Zhelang and Ma, Yizhuo and Huang, Xiaowen and Meng, Jintao, REXIO: Indexing for Low Write Amplification by Reducing Extra I/Os in Key-Value Store Under Mixed Read/Write Workloads,International Conference on Web Information Systems Engineering,245--258,2024
Download Paper | Download Slides