A Fast File System for UNIX*

Review by James Ezick and Ken Hopkinson, Feb 1998

Goals

• Speculate about the nature of file system use and extrapolate possible improvements
• Present performance and efficiency enhancements to the existing UNIX file system.
• Compare and contrast performance and efficiency benchmarks after implementation of enhancements.

Modifications

• Increasing the block size (from 512 bytes to >= 4096 bytes / block)
• Using principles of locality to position data block on the disk thereby reducing latency, etc.
• Long filenames
• File Locking (replacing the need for conventional "lock files")
• Symbolic Links (both relative and absolute path names)
• Increase the reliability of rename (without copying)
• Quotas

Implementation

• 4096 or larger blocks 1024 block segments.
• Scatter the superblock across multiple clusters.
• Superblock contained a vector os lists called rotational layout tables.
• Bitmap array to hold "free" segments.
• Three step allocation hierarchy for writing files that had increased in size.
• Cylinder groups comprised of one or more consecutive cylinders on a disk, used to organize files within a directory.
• Global allocator call the local allocator which uses a four level allocation strategy for data placement.

Achievements

• Intelligent organization of directories and files to take advantage of locality resulting in increased throughput.
• Developed a system to derive the benefits of large blocks while still coping with the resulting waste of internal fragmentation.
• Made the existing file system more robust through the addition of useful features.

Drawbacks

• High CPU usage is required to implements solutions to problems resulting from larger block size (impractical for commercial use).
• The number of small files that tend to exist on a file system prevent larger blocks from being feasible without some way of dealing with internal fragmentation (intolerable amount of waste).
• Quotas do not deal with files created by a user, hard linked to by another user, and deleted by the creator.
• No built in facility for eliminating "dangling" symbolic links.
• File system relies on a machine dependent parameter that can cause a performance drop if the file system is moved between physical machines.

Questions

• Are the assumptions on which these improvements are based still valid today.  (I.e., Do most file systems still contain many short files?  Do larger blocks still increase performance?)
• Given the power of today's CPU's, would the percentage of cycles needed to implement the proposed improvements be the about the same?  Is the cost still prohibitive?
• Which optimizations still make sense given today's disks?
• Do you know the parameters necessary to optimize the disk drive in your workstation?
• What do you think of the "enhancements"? Symlinks, locks, long file names, rename, quotas.

* UNIX was a trademark of Bell Laboratories at the time...