面對如今大數據時代的來臨，資料所需的儲存空間比起之前可說是指數般的增加，但現今的磁碟容量密度已到了一個表面密度的極限時，我們採用了疊瓦式硬碟(Shingled Magnetic Recording Disk)，而疊瓦式硬碟顧名思義就是，本身是因為利用類似屋瓦層疊的方式去排列其磁軌，故使其能夠增加磁錄密度，但其相對於傳統硬碟，亦有一些限制，而這限制即是在做隨機寫入操作時會需要以 Read-Merge-Write 的方式去寫入以確保其磁碟中資料的正確性，但這樣的方式卻為疊瓦式硬碟 (Shingled Magnetic Recording Disk) 帶來了額外的寫入負擔，而這就是所謂的寫入放大 (Write Amplification) ，而這樣的限制即是SMR與傳統硬碟相比起來最大的限制。
而我的研究就是將B+-tree這個檔案系統中與資料庫系統中常用的定址結構應用於新型態的疊瓦式硬碟環境中，但因B+-tree這樣的資料結構，在疊瓦式硬碟應用的情景中會產生許多效能上的額外的負擔，而我們所設計出的系統就是因應這樣效能的問題，並配合設計的元件使其能夠有效的減輕此資料結構在疊瓦式硬碟環境中運作上效能的負擔，同時兼顧有疊瓦式硬碟 (Shingled Magnetic Recording Disk) 大容量的特性。

Facing the big data era, traditional magnetic hard disk drives (HDDs) can no longer hold the capacity requirement of such big data applications. Due to traditional magnetic hard disk drives is going to reach its areal density limitation. Shingled magnetic recording (SMR) is a promising solution that SMR overlaps the tracks in a similar fashion to shingle on the roof of a house. Although shingled magnetic recording has the big capacity this good property, it still has some limitation. For random write operation, we need do it with a read-merge-write (RMW) process to ensure the data correct. It is RMW process that caused the write amplification problem on SMR.
The B+-tree is a common data structure often used in the file system and database system as index system because it has some good feature. It enables the Filesystem and Database system to find the leaf node quickly and always in a stable time. And B+-tree is a self-balancing tree, it will never appear skew scenario. But we find that this kind data structure may have huge operation overhead on shingled magnetic recording drives. In our research, we try to design a system to mitigate B+-tree overhead on shingled magnetic recording disk and leverage the SMR big capacity in the meanwhile.
In this work, we considered B+-tree on the host-managed SMR drives. We used CPU and main memory as auxiliary resource to enhance B+-tree performance on SMR drive. Whole methodology in our work include two parts elementary data structure and garbage collection policy.

1.介紹 1
2.背景與研究動機 5
3.研究方法 9
3.1 方法概覽 9
3.2 基礎資料結構 9
3.2.1 Outplace update與節點映射表 9
3.2.2 分離葉節點與內部節點於不同磁區 10
3.2.3 在write pointer直接做 inplace update 10
3.3 Garbage Collection策略 11
3.3.1 Garbage Collection過程中的輔助資料 11
3.3.2 Garbage Collection 演算法 12
3.3.3 清理節點映射表(Node Mapping Table) 13
4. 實驗 15
4.1 實驗設置 15
4.2 實驗結果 16
5. 結論 17
6. 附錄 18

[1] Another layer of indirection @ONLINE, http://hselin.com/blog/2015/08/
26/another-layer-of-indirection/, 2015.
[2] Abutalib Aghayev and Peter Desnoyers. Skylight—a window on
shingled disk operation. In 13th USENIX Conference on File and Storage
Technologies, pages 135–149, 2015.
[3] A. Amer, J. Holliday, D. D. E. Long, E. L. Miller, J.-F. Pris, and T. Schwarz. Data management and layout for shingled magnetic recording. In IEEE Transactions on Magnetics, October 2011.