|
[20080904]
|
Speeding up operations with directory hashing
While working on UDF, Reinoud Zandijk has added routines
to speed up directory operations. Remember that in FFS,
directories are list of
struct dirent
entries that are stored one after another. This makes lookup sequential,
deleting leaves holes, and inserting needs to check for possible
holes. In other words: it's suboptimal.
The proper solution would be to move some tree-based directory
format, but that's not around the corner - file systems like
SGI's XFS have that,
and anyone's welcome to do the work on FFS.
Another option to mitigate the suboptimal on-disk format is
to use in-memory caching and fast lookup, which is just what
Reinoud did. His runs of the "postmark" benchmark show some
impressive speedups:
| What | Old | New
| | File create: | 2/s | 208/s
| | File create alone: | 4/s | 230/s
| | File create + transaction: | 1/s | 318/s
| | File read: | 1/s | 330/s
| | File append: | 1/s | 336/s
| | File delete: | 2/s | 208/s
| | File delete alone: | 9/s | 970/s
| | File delete + transaction: | 1/s | 348/s
| | Data read: | 3 kb/s | 283kb/s
| | Data written: | 14kb/s | 1150kb/s
|
See
the commit log
for more information.
P.S.: Someone please add wikipedia pages for FFS/UFS!
P.P.S.: Yes, this has been in UFS/FFS as "dirhash" for some time
[Tags: udf]
|
