Bench
Numbers from my own shelf, on my own files. They are not a review of anything; a different set of files would reorder them.
Restore of 41,000 small files (7.4 GB)
| Target | Wall clock | Sequential claim |
|---|---|---|
| Mirrored pair, 7200 rpm | 18 min 40 s | 170 MB/s |
| Single 5400 rpm drive | 26 min 05 s | 140 MB/s |
| SATA SSD, three years old | 3 min 12 s | 510 MB/s |
| USB enclosure, same SSD | 6 min 55 s | 430 MB/s |
The enclosure loses half its advantage on small files while advertising almost the same sequential figure. That gap is the whole point of measuring the thing you actually do.
Method
Drop caches, copy with rsync -a from a warm source, time the whole
run including the final sync. Three runs, middle value recorded. Room
at 21 °C, nothing else running, no compression anywhere in the path.
What I stopped measuring
Peak throughput, queue depths above four, and anything reported by the drive's own firmware. None of them predicted a restore time on this shelf.