But is the drive's reordering buffer so deep? NCQ in the system reorders only up to 32 requests.
I doubt the drive's firmware is capable of reordering millions of requests. These things have typically small amount of memory, just enough for typical use-cases.
Keeping a small window for sorting a subset of requests is much weaker than sorting all of them up front. Even if the window is 50% of all files, that would mean the average distance between requests increases 2 times (assuming blocks are distributed uniformly).
Keeping a small window for sorting a subset of requests is much weaker than sorting all of them up front. Even if the window is 50% of all files, that would mean the average distance between requests increases 2 times (assuming blocks are distributed uniformly).