1. The Design and Implementation of a Log-Structured File System
Rosenblum, Mendel, and John K. Ousterhout, ACM Transactions on Computer Systems (TOCS) 10.1 (1992):
Presentation by Sion Lee

2. Introduction
Designs for file systems of the 1990’s
Log-structured file systems
Crash recovery
Experience with the Sprite LFS
Conclusion

3. Log-structured file system
1. Introduction
Log-structured file system
“Log”
Crash recovery
Segment
Segment cleaner
Unix FFS vs Sprite LFS
<Ext2 FS>
<Sprite LFS>

4. 1. Introduction
References

5. 2. Designs for file systems of the 1990’s
Technology
CPU
Speed↑
Main memory
Size ↑  cache size↑
Disk
Cost
Capacity
Performance  transfer bandwidth, access time
Workload
Small-file access

6. 2. Designs for file systems of the 1990’s
Problem?
1) Spread information  frequent access
2) Inode is separate from the file
3) Synchronous write

7. 3. Log-structured file systems
Improve write performance  buffering & writing at once
Issue
Retrieve information
Manage free space

8. 3. Log-structured file systems
Inode
Write log sequentially
Mapping with inode map
Includes metadata (indirect block address, etc)
Inode map
Its address resides in cp
Mapping inode number with actual address of inode

9. 3. Log-structured file systems
Segment
512KB or 1MB
Include 4KB blocks
Threading, copying
Segment cleaning
1) Read a number of segments into memory
2) Identify the live data
3) Write the live data back to a smaller number of clean segments
4) Segments that were read are marked as clean

10. 3. Log-structured file systems
Segment summary block
Useful during crash recovery
Distinguish live blocks from those that have been overwritten or deleted
Records the block age
Its address resides in cp
Identify alive blocks, file #, block #, etc
uid = version # + inode #
Segment usage table
For segment cleaning (cost-benefit)
# of live bytes, recent modified time

11. 3. Log-structured file systems
Segment cleaning policy
1) When should the segment cleaner execute?
2) How many segments should it clean at a time?
3) Which segments should be cleaned?
4) How should the live blocks be grouped when they are written out?

12. 3. Log-structured file systems

13. 3. Log-structured file systems

14. 3. Log-structured file systems

15. Traditional Unix file systems  fsck (high cost)
4. Crash recovery
Traditional Unix file systems  fsck (high cost)
LFS  last location of log
Checkpoint
Roll-forward
<Window system crash>

16. Checkpoint Fixed position on disk
4. Crash recovery
Checkpoint
Fixed position on disk
Addresses of inode map, segment usage table, current time, pointer to the last segment written
Two checkpoint
One for system crash
The other for crash during checkpoint operation
Do checkpoint when fs is unmounted or system shut down

17. 4. Crash recovery
Roll-forward
Scan through the log segments that were written after the last checkpoint
Segment summary block
Adjust the utilization in the segment usage table read from the checkpoint
Restore consistency between directory entries and inodes
Directory operation log
Log →
new segments
∙∙∙
Last checkpoint segment

18. 5. Experience with the Sprite LFS
Sprite LFS is faster than SunOS.
Sprite LFS only kept the disk 17% busy during the create phase while saturating the CPU.
Sprite LFS has a higher write bandwidth and the same read bandwidth as SunOS.
Traditional file system: logical locality
Log-structured file system: temporal locality

19. 5. Experience with the Sprite LFS
Cleaning costs are lower in Sprite LFS than in the simulations.
1) all the files in the simulations were just a single block long
2) simulated reference patterns were evenly distributed within the hot and cold file groups

20. 5. Experience with the Sprite LFS

21. 6. Conclusion
LFS
Collect large amounts of new data in a cache (main memory)
Write the data to disk in a single large I/O that can use all of the disk’s bandwidth
Low cleaning overheads can be achieved  cost and benefit
 Use disks an order of magnitude more efficiently than existing file systems

22. 2019.09.23 Presentation by Sion Lee sioni322@naver.com
Thank you
Presentation by Sion Lee