Thats what it is. The specific "why" is there aren't many O(1) directory traverses.
So a million files in a dir is going to take longer to access any individual file than if there's only 3 files to pick from.
And if it scales worse than linear, a tree structure, although hitting the FS multiple times once at each level, in total can take less time.
Finally if you can avoid a smooth distribution hash and intentionally order by something important (time?) then you only need a cache the most recent directories in memory and the deep historical archive can fend for itself rarely accessed without getting in the way of the busy files. If you rarely if ever leave /stuff/thisYear/today/ then whatever is in /stuff/2011/dec25 will never slow today down or get in the way.
So a million files in a dir is going to take longer to access any individual file than if there's only 3 files to pick from.
And if it scales worse than linear, a tree structure, although hitting the FS multiple times once at each level, in total can take less time.
Finally if you can avoid a smooth distribution hash and intentionally order by something important (time?) then you only need a cache the most recent directories in memory and the deep historical archive can fend for itself rarely accessed without getting in the way of the busy files. If you rarely if ever leave /stuff/thisYear/today/ then whatever is in /stuff/2011/dec25 will never slow today down or get in the way.