非揮發性記憶體 ( NVM ) 不僅在系統斷電之後保證資料不會消失，還提供了和DRAM相同等級的讀寫速度。而基於DRAM-SSD的日誌結構合併樹 (LSM-Tree)是一種分層的文件管理方法，被廣泛應用於鍵-值的存儲系統，可以利用混和NVM的架構來取得更好的性能。先前已經有許多這方面的研究了，但其中大多數都沒有考慮到鍵的實際使用情況。因此，在此篇論文中，我們提出了 TbLSM，一種在DRAM-NVM-SSD架構上基於鍵溫度的數據壓縮方法的LSM-Tree，除了能減少 DRAM 和 SSD之間的序列化-反序列化開銷，還通過永久存儲的分層設計來縮短被頻繁訪問的鍵其讀取延遲。設計核心在於: TbLSM是利用了 NVM 的字節尋址能力 ( byte-addressability ) 和低延遲 ( low latency ) 的特性，把鍵按照其訪問頻率來進行分類。我們將 TbLSM 與以下基於 LSM-Tree 的系統進行比較：TLSM 和 LevelDB-NVM。經過評估表明，使用基於溫度的壓縮的 TbLSM 能夠比 TLSM 和 LevelDB-NVM 具有更好的性能。

Non-volatile memory (NVM) not only provides data persistence but is almost as fast as DRAM. The Log-Structured Merge-tree (LSM-Tree) is a file management data structure, which is based on the DRAM-SSD architecture and is widely used in key-value storage systems. It can achieve better performance by using the hybrid architecture with NVM. There's a lot of research on this, but most of them do not take into account the key usage. In this paper, we proposed TbLSM, temperature-based data compaction for LSM-tree with NVM to reduce the overhead of serialization-deserialization between DRAM and SSD and to shorten the read latency of frequent-accessed keys by the persistent storage tiering design. The core idea is that: TbLSM leverages the byte-addressability and low latency of NVM to classify keys by their access frequency. We compare TbLSM with LSM-Tree-based systems: TLSM and LevelDB-NVM. Our evaluations show that TbLSM which uses temperature-based compaction has better performance than TLSM and LevelDB-NVM.

Contents
Abstract (Chinese) I
Abstract II
Contents III
List of Figures V
List of Tables VI
1 Introduction 2
2 Background & Related work 4
2.1 LSM-tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.1.1 LevelDB . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 NVM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2.1 PMDK . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3 LSM-tree with NVM . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.3.1 NoveLSM . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.3.2 TLSM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
3 Motivation 10
4 Methodology 11
III
4.1 Design of Node Structure . . . . . . . . . . . . . . . . . . . . . . . . 11
4.2 Temperature-based Compaction . . . . . . . . . . . . . . . . . . . . 12
4.2.1 Level definition . . . . . . . . . . . . . . . . . . . . . . . . . 12
4.2.2 Trigger Compaction . . . . . . . . . . . . . . . . . . . . . . . 14
4.2.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
4.2.4 Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
5 Evaluation 20
5.1 Experimental Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
5.2 Workloads . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
5.2.1 DB Bench . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
5.2.2 YCSB . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
5.3 Set Reference Times Threshold . . . . . . . . . . . . . . . . . . . . 22
5.4 Overhead And Performance . . . . . . . . . . . . . . . . . . . . . . 23
5.4.1 Overhead . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
5.4.2 Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
5.5 The Impact Of Data Size . . . . . . . . . . . . . . . . . . . . . . . . 27
6 Conclusion 29
Bibliography 30

[1] Joy Arulraj and Andrew Pavlo. How to build a non-volatile memory database
management system. In Proceedings of the 2017 ACM International Confer-
ence on Management of Data, pages 1753–1758, 2017.
[2] Fay Chang, Jeffrey Dean, Sanjay Ghemawat, Wilson C Hsieh, Deborah A Wal-
lach, Mike Burrows, Tushar Chandra, Andrew Fikes, and Robert E Gruber.
Bigtable: A distributed storage system for structured data. ACM Transac-
tions on Computer Systems (TOCS), 26(2):1–26, 2008.
[3] Sanjay Ghemawat and Jeff Dean. Leveldb. https://github.com/google/
leveldb.
[4] Sudarsun Kannan, Nitish Bhat, Ada Gavrilovska, Andrea Arpaci-Dusseau,
and Remzi Arpaci-Dusseau. Redesigning {LSMs} for nonvolatile memory
with {NoveLSM}. In 2018 USENIX Annual Technical Conference (USENIX
ATC 18), pages 993–1005, 2018.
[5] Jihwan Lee, Won Gi Choi, Doyoung Kim, Hanseung Sung, and Sanghyun
Park. Tlsm github open source. https://github.com/hwani3142/
leveldb-pmdk.
[6] Jihwan Lee, Won Gi Choi, Doyoung Kim, Hanseung Sung, and Sanghyun
Park. Tlsm: Tiered log-structured merge-tree utilizing non-volatile memory.
IEEE Access, 8:100948–100962, 2020.
30
[7] Wenjie Li, Dejun Jiang, Jin Xiong, and Yungang Bao. Hilsm: an lsm-based
key-value store for hybrid nvm-ssd storage systems. In Proceedings of the
17th ACM International Conference on Computing Frontiers, pages 208–216,
2020.
[8] Apache HBase team. Apache hbase. https://hbase.apache.org/.
[9] Cassandra team. Cassandra. http://cassandra.apache.org/.
[10] RocksDB team. Rocksdb. https://github.com/facebook/rocksdb.
[11] Wikipedia contributors. Cold data — Wikipedia, the free encyclope-
dia. https://en.wikipedia.org/w/index.php?title=Cold_data&oldid=
1094308249, 2022. [Online; accessed 22-September-2022].
[12] Wikipedia contributors. Leveldb — Wikipedia, the free encyclope-
dia. https://en.wikipedia.org/w/index.php?title=LevelDB&oldid=
1108873612, 2022. [Online; accessed 19-September-2022].
[13] Ling Zhan, Kai Lu, Zhilong Cheng, and Jiguang Wan. Rangekv: an effi-
cient key-value store based on hybrid dram-nvm-ssd storage structure. IEEE
Access, 8:154518–154529, 2020.