Skip to content

Latest commit

 

History

History
54 lines (33 loc) · 1.79 KB

[FAST21]ROART.md

File metadata and controls

54 lines (33 loc) · 1.79 KB

Title: ROART: Range-query Optimized Persistent ART

Source: FAST'21

Authors: Shaonan Ma, Kang Chen, Shimin Chen, Mengxing Liu, Jianglang Zhu, Hongbo Kang, and Yongwei Wu.


Summary

  • The problem the paper aims to solve

    • B^+^-Tree and radix tree do not support variable-sized keys and have relative low range queries performance.
    • Current allocators can be improved to adopt the poor write scalability in persistent memory.
  • How can the paper address the problem? What is the main idea of this paper?

    • Propose Range queries support based on ART
      • Make leaf node support variable-sized keys
    • Optimize ROART
      • Leaf compaction
      • Entry compression
    • Selectively metadata persistence
      • Minimally ordered split
    • DCMM
      • Propose a new allocator to improve performance
      • Enable instant restart and fast recovery
  • Validations

    Their evaluations show that ROART outperforms the stat-of-the-art B^+^-Trees by 1.17∼8.27X respectively.


Strengthens

  • ROART enables range query support while achieves better performance compared with B^+^-Tree. This could enlighten complicated optimization in future.
  • This paper proposes a delayed check memory management which enables fast recovery and instant restart. At the same time, it can outperform other allocators greatly. It will show the great impact in the future.

Weaknesses

  • The overall idea in this paper is normal and everywhere, which makes it less innovative.
  • The paper writing shades too many details, making it hard to follow and understand.
  • Evaluations do not compare the time used by recovery with other methods.

Comments

  • This work is highly incremental and lack of innovation. I would reject this paper very weakly if I were a reviewer.