Deduplication Triggered Compaction For Lsm-Tree Based Key-Value Store

Weitao Zhang,Yinlong Xu

DOI: https://doi.org/10.1109/icsess.2018.8663897

2018-01-01

Abstract:Key-value store is a type of NoSQL database providing much higher read and write throughput than traditional SQL databases, LSM-tree is a widely used data structure to improve the Herformance of small write in KV Stores. But its compaction induce to large write amplification. So tiered LSM-tree, which divides each level in LSM-tree implementation into multiple sublevels to reduce tile write amplification. Nowadays, the KV items are frequently updated in many applications, which induces too many stale copies of the same KV items in KV stores, i.e., space amplification. To achieve both low write amplification and low space amplification at the same time, we propose a novel compaction algorithm for tiered LSM-tree and we implemented it on FlameDB, a key-value store. Our experiments show that FlanieDB achieves almost the same low space amplification with LeveiDB and tile comparably low write amplification with LSMtrie.