在現代的電子產品中，儲存記憶體扮演著不可或缺的角色，而其中非揮發性隨機存取記憶體（NVRAM）有著巨大的潛力可以成為未來主流的儲存媒介。NVRAM有著耐震、省電、高效能與延展性等特點，並且儲存的資料即使切斷電源仍然不會消失，再重新供電後可以直接讀取數據的特色，是個具有發展性的儲存媒介，此外Multi-Level-Cell（MLC）NVRAM的引進也讓單位儲存量有了飛躍性的發展。同時，根據我們的觀察，目前常見的Journaling File System(JFS)日誌檔案系統上，儲存資料空間分為兩個區塊，包括Journaling區塊與data區塊，而為了確保數據的完整性，檔案系統會在Commitment時將資料寫入Journaling區塊一次並製作一個transaction來對資料做追蹤，而只有當完整的transaction通過檢查點Checkpoint才會被視為一個有效的一筆transaction。最後通過Checkpoint的Transaction才會被寫入data區作資料的保存。但這個做法將會在Commitment與Checkpoint中會各寫入一次一樣的data，造成同樣的資料重複寫入，這樣的行為對有限耐久度的NVRAM來說是非常不合理與低效率的做法。加上，MLC NVRAM也存在著較低的耐久度與較長的寫入延遲等問題存在，這些缺點會對使用NVRAM的系統造成負面的影響。有鑑於此本篇論文提出一個多重寫入模式的日誌檔案系統(mwJFS)，利用通過對不同的數據提供不同的資料保存時間，並依照NVRAM的優點採用合適的寫入模式來降低其缺點所帶來的負面影響，在保留JFS性能的同時做改進與提高效能。
    根據近年來的研究顯示，我們可以在MLC NVRAM上對應不同的數據保留時間採取不同的寫入操作，例如可以使用Heavy Write寫入使得數據得以保存更長的時間或者使用Light Write讓數據保留時間雖然較Heavy Write來的短，但同時換來寫入時間相對較短，並對NVRAM的損耗也較小的好處。而本文多重寫入模式的日誌檔案系統(mwJFS)使用了四級數據保留機制，此功能用處在於對不同生命周期的數據要求做分類並分為四個級別，分別為分鐘、小時、天與月。另一方面mwJFS包含了一個可以檢查數據是否可以被壓縮的檢查器，用意在於分辨適合的數據保留時間。在配合使用Heavy Write與Light Write的環境下為數據的保留時間做正確的估測與分配。而最後一項機制為數據保留時間監測機制，它將跟蹤所有分類的數據剩餘壽命，而當出現資料所需的保留時間較預期來的長之情況發生時，它將重新分配與刷新資料的保留時間。
    本篇論文透過以上方法達到為數據做適合的保留時間分配，並減少多餘的寫入能量與時間的效果。期望在新型態的記憶體出現時，能適時提供一個量身訂做Journaling File System模型，而透過實驗我們也證實提出的新功能將減少NVRAM的負擔，並且在不影響Journaling File System功能的情形下改善效能等問題。

In recent years, Non-volatile random-access memory (NVRAM) is regarded as a great alternative storage medium because of its beneficial features, including byte addressability, non-volatility, and short read/write latency. In addition, multi-level-cell (MLC) NVRAM was proposed to provide higher storage density.
Meanwhile, in the conventional Journaling File System (JFS), the storage space is divided into two regions, i.e. journal region and data region. To ensure data integrity, the data of JFS are written into journal region first during the commit operation. Data written to the journal region will be enclosed and tracked by a transaction structure, which records both the modified metadata and data content. And the transaction is regarded as a valid one only if it contains both the transaction header (H) and transaction tail (T) to indicate the start and the end points on storage. Then the Transaction pass the Checkpoint operation will be written to the data area. Hence, the same data will be written into storage twice in the Commitment operation and Checkpoint operation. Although JFS enhances the data safety, it also induces the data amplification problem. To mitigate the negative impact of the lengthened write latency of MLC NVRAM on existing JFS, we presents the journaling file system with multi-write mode (mwJFS) to boost the access performance of MLC NVRAM-based JFS.
The proposed mwJFS introduces a four-level data retention model to satisfy the retention requirement of different data types and applies suitable write modes for each retention level. The retention levels are defined in four levels: minute, hour, day and month. The system will take different write operations on MLC NVRAM for different data retention times. For example, the proposed method can use Heavy Write to make data save for a longer period or use Light Write to make data retention time and write Latency shorter than Heavy Write. In addition, a compressible check model is included to classify file data. A retention monitor mechanism and an incremental retention strategy are also included to check the residual data lifetime and refresh when necessary.
The experimental results show that the proposed system can differentiate the data retention requirement of journaled data and applies different write modes to enhance the access performance with lower energy consumption effectively. The mwJFS can reduce the burden of NVRAM and improve overall performance.

目錄
致謝           i
中文摘要       ii
英文摘要       iv
目錄           vi
圖目錄       VIII
表目錄         IX
第一章	緒論	1
1.1	前言	1
1.2	動機與目的	2
1.3	論文章節架構	3
第二章	背景知識與相關文獻	4
2.1非揮發性記憶體（Non-Volatile Memory）	4
2.1.1相變化記憶體(Phase Change Memory)	5
2.1.2 多層單元MLC (Multi-Level Cell)	7
2.2 檔案系統	8
2.2.1  EXT3檔案系統	9
2.2.2  日誌檔案系統（Journaling File System）	10
第三章	Multi-Write Modes JFS	12
3.1	系統架構與整體流程	13
3.2	資料保留時間分類機制	15
3.3	四種等級的保留時間與In-place、Out-place	18
3.4	Incremental Retention Strategy	21
3.5	總體流程	26
第四章	研究成果	28
4.1實驗環境	28
4.2 模擬I- IOzone	32
4.2.1 	IOzone - Energy 比較分析	35
4.2.2   IOzone - Latency 比較分析	39
4.3 模擬II – Postmark	41
4.3.1 	Postmark - Energy 比較分析	44
4.3.2 	Postmark - Latency比較分析	46
第五章	結論與未來展望	48
參考文獻	49

圖目錄

圖2.1  各種記憶體的寫入速度(4GHz processor cycles)[2]	5
圖2.2  PCM的工作原理[2]	6
圖2.3  MLC技術[4]	8
圖2.4  Journaling File System機制	11
圖3.1  整體系統架構	13
圖3.2  系統架構運作流程	14
圖3.3  不同種類的資料分類	16
圖3.4  四種等級的資料保存時間與各級資料分配	18
圖3.5  Out-place與In-place	20
圖3.6  在Transaction中新增新的INFO	22
圖3.7  Incremental Retention Strategy	23
圖3.8  Refresh機制	24
圖3.9  整體流程圖	26
圖4.1  IOzone Energy之比較結果	38
圖4.2  IOzone Latency之比較結果	40
圖4.3  Postmark Energy之比較結果	45
圖4.4  Postmark Latency之比較結果	47

表目錄

表2.1  各儲存媒介的性能比較表[2]	7
表3.1	STT-RAM不同的數據保留時間[9]	15
表4.1	模擬器規格與實驗數據[17][18]	29
表4.2	因資料保留等級不同所能提升之效能[18]	31
表4.3	不同保留等級下之寫入能量與寫入時間	32
表4.4	IOzone之EXT3模擬數據	33
表4.5	IOzone之mwJFS模擬IO寫入量	34
表4.6	IOzone之mwJFS Refresh Overhead寫入數據	35
表4.7	Energy(write only)比較模擬數據(單位: pJ)	36
表4.8	Energy(write only)比較模擬結果	37
表4.9	Latency(write only)比較模擬數據(單位: μs)	39
表4.10  Latency(write only)比較結果	40
表4.11  Postmark之EXT3模擬數據	41
表4.12  Postmark之mwJFS 模擬IO寫入量	42
表4.13  Postmark之mwJFS Refresh Overhead寫入數據	43
表4.14  Energy(write only)比較模擬數據(單位:pJ)	44
表4.15  Energy(write only)比較模擬結果	45
表4.16  Latency(write only)比較模擬數據(單位:μs)	46
表4.17  Latency(write only)比較結果	47

參考文獻
[1]	Storage Class Memory: What’s Next in Enterprise Storage. Sundar Ranganathan .https://blog.netapp.com/storage-class-memory-what-is-next-in-enterprise-storage/
[2]	Moinuddin K. Qureshi Vijayalakshmi Srinivasan Jude A. Rivers. Scalable High Performance Main Memory System Using Phase-Change Memory Technology. ISCA’09, June 20–24, 2009
[3]	W. Zhang and T. Li. 2011. Helmet: A resistance drift resilient architecture for multi-levelcell phase change memory system. In 2011 IEEE/IFIP 41st International Conference on Dependable Systems Networks (DSN).
[4]	C. Pan, M. Xie, J. Hu, Y. Chen, and C. Yang. 2014. 3M-PCM: Exploiting Multiple Write Modes MLC Phase Change Main Memory in Embedded Systems. In Proceedings of the2014 International Conference on Hardware/Software Codesign and System Synthesis.
[5]	A. Jog, A. K. Mishra, C. Xu, Y. Xie, V. Narayanan, R. Iyer, and C. R. Das. 2012. Cache revive: Architecting volatile STT-RAM caches for enhanced performance in CMPs. In DAC Design Automation Conference 2012.
[6]	File System Foresic Analysis.brian carrier.March 17,2005.ISBN:0-32-126817-2
[7]	R. H. Arpaci-Dusseau and A. C. Arpaci-Dusseau. 2015. Operating Systems: Three Easy Pieces, chapter Crash Consistency: FSCK and Journaling. Arpaci-Dusseau Books (2015).
[8]	C. Chen, J. Yang, Q. Wei, C. Wang, and M. Xue. 2016. Fine-grained metadata journaling on NVM. In 2016 32nd Symposium on Mass Storage Systems and Technologies (MSST).
[9]	Mohammad Hossein Samavatian, Hamed Abbasitabar, Mohammad Arjomand, and Hamid Sarbazi-Azad, An Efficient STT-RAM Last Level Cache Architecture for GPUs. 2014 51st ACM/EDAC/IEEE Design Automation Conference (DAC)
[10]	Clinton W. Smullen, Vidyabhushan Mohan. Relaxing Non- Volatility for Fast and Energy-Efficient STT-RAMCaches. 2011 IEEE 17th International Symposium on High Performance Computer Architecture
[11]	Jawad B. Khan. 2013. Method to detect uncompressible data in mass storage device. (2013-01-03). https://www.google.com/patents/US20130007346
[12]	T. Y. Chen, Y. H. Chang, S. H. Chen, C. C. Kuo, M. C. Yang, H. W. Wei, and W. K. Shih.Enabling Write-Reduction Strategy for Journaling File Systems over Byte-addressable NVRAM. In Proceedings of the 54th Annual Design Automation Conference 2017.
[13]	E. Lee, S. Hoon Yoo, and H. Bahn. 2015. Design and Implementation of a Journaling File System for Phase-Change Memory. IEEE Trans. Comput. (2015).
[14]	Professional Linux® Kernel Architecture. Wolfgang Mauerer. ISBN：978-0-470-34343-2
[15]	IOzone. 2016. IOzone Filesystem Benchmark. http://www.iozone.org/. (2016).
[16]	Jeffrey Katcher. 1997. PostMark: a new file system benchmark. Network Appliance Tech Report TR3022. (Oct 1997).
[17]	Y. Park and J. s. Kim. 2011. zFTL: power-effcient data compression support for NAND flash-based consumer electronics devices. IEEE Transactions on Consumer Electronics(2011).
[18]	R. S. Liu, D. Y. Shen, C. L. Yang, S. C. Yu, and C. Y. MichaelWang. 2014. NVM Duet: Unified Working Memory and Persistent Store Architecture. SIGPLAN Not. (2014), 455–470.
[19]	Redaelli, A. Pirovano, A. Locatelli, and F. Pellizzer. 2008. Numerical Implementation of Low Field Resistance Drift for Phase Change Memory Simulations. In 2008 Joint Non-Volatile Semiconductor Memory Workshop and International Conference on Memory Technology and Design.
[20]	Z. Zhang, L. Ju, and Z. Jia. 2016. Unified DRAM and NVM hybrid buffer cache architecture for reducing journaling overhead. In 2016 Design, Automation Test in Europe Conference Exhibition (DATE).