On Wed, Feb 15, 2023 at 03:10:08PM +0100, Why 42? The lists account. wrote: > However, I also tried testing the same two filesystems using the > "Flexible IO Tester" or fio (it's available as a package). When I used it > to do random 4K reads and writes, I appear to have the opposite result:
... > I wonder why that would be? For a start, I would test using something other than /dev/zero as the data source. It's entirely possible that the firmware on an SSD would special case writing a block that contains only 0x00 bytes. In that case, and assuming that the filesystem block boundaries align with the SSD's own internal flash block layout, the SSD would only need to update it's metadata to point those LBA blocks to an internal 'zero' block. This would virtually eliminate the overhead of actually writing to the flash, and allow it to accept data from the host at a much faster speed. As soon as you write a single non-0x00 byte, the drive would have to do a propper write to the main flash memory and not just the area which contains it's internal LBA to flash block mapping, (which may also be write-cached). Depending on the state of the SSD, (recently secerased, used with another OS which supports TRIM, alignment of the filesystem blocks with the raw flash blocks, etc, etc), this could mean either a write, or a read-erase-write cycle. Using /dev/zero as the source definitely makes it a synthetic benchmark.
