1. Embedded System Lab. 정영진 cyj4369@gmail.com The Design and Implementation of a Log-Structured File System Mendel Rosenblum and John K. Ousterhout ACM Transactions on Computer Systems, Vol. 10, Issue 1, pp. 26~52, 1992

2. 정 영 진 Embedded System Lab.

3. 정 영 진 Embedded System Lab. Introduction Log-structured file system Simulation Crash recovery Evaluation Conclusion Contents

4. 정 영 진 Embedded System Lab. Introduction A gap between processor and HDD http://blogs.msdn.com/b/e7/archive/2009/01/25/disk-defragmentation-background-and-engineering-the-windows-7-improvements.aspx

5. 정 영 진 Embedded System Lab. Internal structure of HDD  Two components of disk performance Transfer bandwidth Access time Introduction Chris Ruemmler and John Wilkes. An introduction to disk drive modeling. IEEE Computer, 27(3):17–28, March 1994.

6. 정 영 진 Embedded System Lab. Workloads  Small file Office, engineering applications, etc … ‘Metadata’  Large file Limited by the bandwidth of the I/O Limited by memory subsystems  LFS Focus on the efficiency of small file Introduction

7. 정 영 진 Embedded System Lab. Problem with existing file systems  Example of current file system Unix FFS  Spread information Cause too many small access  Synchronous write File system metadata Introduction request write request write request write

8. 정 영 진 Embedded System Lab. Fundamental idea  Asynchronous write  Writing all the changes to disk sequentially File data, attributes, index blocks, directories, … Single operation Log-structured file system request Buffer write

9. 정 영 진 Embedded System Lab. LFS(Log-structured file system)  LFS stores data permanently in the log  The log contains indexing information  Large extents of free space available for writing new data Segment Segment cleaner Cleaning policy Log-structured file system

10. 정 영 진 Embedded System Lab. LFS layout  512KB ~ 1MB Segment Log-structured file system http://home.postech.ac.kr/~kypark/study/fs.html

11. 정 영 진 Embedded System Lab. File location and reading  Basic structure is same with FFS  Not fixed inode inode map Cached in main memory  Compact arrangement Log-structured file system

12. 정 영 진 Embedded System Lab. Segment cleaning  Delete and update Fragmentation  Threading Makes fragmentation  Copying Large cost Log-structured file system

13. 정 영 진 Embedded System Lab. Cleaning policy  When?  How many segment?  Which segment?  How to group live blocks? Log-structured file system not sensitive

14. 정 영 진 Embedded System Lab. Which segment?  Write cost, greedy  u : Utilization(fraction of live data, 0 ≤ u < 1)  Cost is determined by u  Trade off Performance & Cost per usable data Log-structured file system

15. 정 영 진 Embedded System Lab. Access pattern  Uniform Same request pattern  Hot and cold Simulation

16. 정 영 진 Embedded System Lab. Simulation result Simulation

17. 정 영 진 Embedded System Lab. New cleaning policy  Cost-benefit policy Benefit components : free space, space is likely to stay free Simulation

18. 정 영 진 Embedded System Lab. Crash recovery  Check point Time check, inode map, last segment pointer, etc …  Roll forward Use segment summary inode is exist / not exist Crash recovery Check point region(fixed) Crash Roll forward

19. 정 영 진 Embedded System Lab. Micro-benchmarks(Small file) Evaluation

20. 정 영 진 Embedded System Lab. Micro-benchmarks(Large file) Evaluation

21. 정 영 진 Embedded System Lab. Performance is as good as previous system or much better. Not only benefit with many small files but also benefit with large files. Conclusion

22. 정 영 진 Embedded System Lab. Chris Ruemmler and John Wilkes. An introduction to disk drive modeling. IEEE Computer, 27(3):17–28, March 1994. http://blogs.msdn.com/ Reference

23. 정 영 진 Embedded System Lab. Q & A