Back when I still ran Linux on my personal servers I noticed this. I was adding TRIM support to the md drivers before the maintainers finally got around to it.

Back when I still ran Linux on my personal servers I noticed this. I was adding TRIM support to the md drivers before the maintainers finally got around to it. That's when I noticed the balancing code tended to favour one disk. I added round-robin code and experimented with splitting large sequental requests between idle disks.