All filesystems were tested on the same partition with offset 1MiB and size 4GiB, mounted with noatime and discard, except tmpfs. The two btrfs tests additionally tries out autodefrag vs nodatacow, in an attempt to isolate the impact of copy-on-write. I really had hoped to run the sqlite benchmark too, but PTS said it failed to produce a result. More overwhelming than these results, is a real sqlite workload I have that takes tens of minutes on a rotating hdd (and slows down the machine) or a split second on tmpfs.