Jan Kara (jack@xxxxxxx) wrote on Mon, Apr 17, 2023 at 02:15:37PM -03:
> Commit 7e55c60acfbb ("md/raid5: Pivot raid5_make_request()") changed the
> order in which requests for underlying disks are created.
In which version was this first applied?
We've observed a large drop in disk performance since 4.* to the point that
6.1.* is almost unusable for some tasks. For example, we have 2 backup
machines, one using raid6/ext4 and another using zfs. The backup of a third
machine with a filesystem with ~25M inodes takes 6x-8x longer on the one using
raid6/ext4 than on the one using zfs... This is during the rsync phase.
Afterwards it removes old trees and the rm with raid6 takes eons even though
the disks are not at all busy (as shown by sar). Running several rm in parallel
speeds things up a lot, showing the problem is not in the disks.
