The enclosed patch (tested against SVN revision 2104) changes the file diff logic with the following enhancements:
- Files are read no more than twice (versus read again and again for every block size).
- Before performing a rolling checksum each server-side block is first checked (by MD5) at its previous location in the file. In the event the block has not changed or moved, the rolling checksum is skipped.
- Rolling checksums are searched in total-file-coverage order (size times number of blocks) favoring larger blocks in the final recipe.
In my testing these changes improve file diff performance wall time from 2-10x and make the diff process CPU bound (instead of IO bound).
No new dependencies are created by the patch, this is only an algorithmic change. The code passes existing unittests. Additionally, I have tested with my personal data for 3 wks on OS X (i386) without incident.
