LSM-Tree Based Hardware System Design for Compaction Acceleration

EasyChair Preprint 15499

Abstract

With the rapid development of technology and the advent of the era of big data, Log-structured merge tree (LSMtree) based key-value storage systems are widely used in data centers that persistently store ultra-large-scale data. This paper proposes decentralizing the compaction operation to FPGA for acceleration, which can free up CPU computational resources and I/O bandwidth, thus improving the quality of service of the overall system. To evaluate the acceleration performance of the compaction module on hardware, this paper also completes the design of the NAND Flash controller and the corresponding storage card. The experimental results show that the compaction module implemented in this work can achieve up to 11.07x acceleration performance compared to LevelDB.

Keyphrases: Compaction, FPGA, LSM-tree, NAND Flash

@booklet{EasyChair:15499,
  author    = {Sirui Peng and Lin Yang and Zhikang Ding and Xingyu Chen and Xu Kong and Haidong Tian and Xiankui Xiong and Xiaoyong Xue and Xiaoyang Zeng},
  title     = {LSM-Tree Based Hardware System Design for Compaction Acceleration},
  howpublished = {EasyChair Preprint 15499},
  year      = {EasyChair, 2024}}