Scholarlink

The mainstream LSM-tree based key-value stores suffer from periodic performance plunges even near zero availability due to write amplification and write stalls. Byte-addressable non-volatile memory, with its low latency and persistence, offers a new opportunity to address these issues. This paper proposes HyLSM, a persistent key-value storage system based on LSM-tree, under a DRAM-NVM-SSD three-tier architecture. HyLSM divides the Memtable into a DRAM Memtable and an NVM Memtable and separates the LSM-tree into hot tiers (stored in NVM) and cold tiers (stored in SSD). Tailored to the characteristics of NVM and SSD, a hybrid compaction mechanism is introduced: fine-grained compaction for NVM tiers and selective compaction for SSD tiers, improving compaction efficiency and alleviating write amplification and write stalls. Through comparative experiments with LevelDB, NVM-based NoveLSM, our optimized version NoveLSM_B, and NVM-based ThanosKV, the results demonstrate that HyLSM achieves significant improvements.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Reducing Write Stall and Write Amplification for LSM-Tree Based Key-Value Stores Using a Hybrid Compression Strategy

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Zhicheng Bu,
  • Jinbao Wang,
  • Lina Chen,
  • Yaozong Yang,
  • Hong Gao

摘要

The mainstream LSM-tree based key-value stores suffer from periodic performance plunges even near zero availability due to write amplification and write stalls. Byte-addressable non-volatile memory, with its low latency and persistence, offers a new opportunity to address these issues. This paper proposes HyLSM, a persistent key-value storage system based on LSM-tree, under a DRAM-NVM-SSD three-tier architecture. HyLSM divides the Memtable into a DRAM Memtable and an NVM Memtable and separates the LSM-tree into hot tiers (stored in NVM) and cold tiers (stored in SSD). Tailored to the characteristics of NVM and SSD, a hybrid compaction mechanism is introduced: fine-grained compaction for NVM tiers and selective compaction for SSD tiers, improving compaction efficiency and alleviating write amplification and write stalls. Through comparative experiments with LevelDB, NVM-based NoveLSM, our optimized version NoveLSM_B, and NVM-based ThanosKV, the results demonstrate that HyLSM achieves significant improvements.