The Design and Implementation of a Log-Structured File System

M. Rosenblum, and J.K. Ousterhout. ACM Transactions on Computer Systems. Vol. 10, No. 1 (Feb 1992), pp. 26-52.

The whole file system is in an append-only log. It's divided into sizable (~1MB) segments (and the log threaded through them) to make cleaning easier without losing performance.

Details:

• Find file locations with inodes; inode map expected to be cached in RAM. Yields comparable read performance.
• They use checkpointing to speed cleaning and recovery
• Multiple cleaning strategies: successful one is to preferentially clean "hot segments" (frequently deleted/overwritten)

Wins:

• Small writes (eliminate metadata overhead)
• Other ops competitive with FFS
• Recovery after crashes is fast
• More of the disk bandwith is used => burden on CPU

Problems:

• Not good on a full disk (> 80% full)
• Relies on write-behind caching -- this would help all FSes
• Not good with a lot of fsyncs (TP)