Hi Thomas, How does your crush map/tree look? If your crush failure domain is by host, then your 96x 8T disks will be as useful as you're 1.6T disks, because smallest failure domain is your limiting factor. So you can either redistribute your disks to be 16x8T+32x1.6T per host, or you could group your 1.6T nodes into groups (chassis perhaps) and move the 8T nodes into their own chassis, and then set your failure domain to chassis, and this would likely lead to a much more even distribution. I imagine right now you're 1.6T disks are nearful, and your 8T disks are anything but. Be careful with something like this however, because you will probably run into some iops discrepancies due to number of spindles/TB difference across 'chassis'. Hope that helps. Reed
On Sep 23, 2019, at 4:07 AM, Thomas <74cmonty@gmail.com> wrote:
Hi,
I'm facing several issues with my ceph cluster (2x MDS, 6x ODS). Here I would like to focus on the issue with pgs backfill_toofull. I assume this is related to the fact that the data distribution on my OSDs is not balanced.
This is the current ceph status: root@ld3955:~# ceph -s cluster: id: 6b1b5117-6e08-4843-93d6-2da3cf8a6bae health: HEALTH_ERR 1 MDSs report slow metadata IOs 78 nearfull osd(s) 1 pool(s) nearfull Reduced data availability: 2 pgs inactive, 2 pgs peering Degraded data redundancy: 304136/153251211 objects degraded (0.198%), 57 pgs degraded, 57 pgs undersized Degraded data redundancy (low space): 265 pgs backfill_toofull 3 pools have too many placement groups 74 slow requests are blocked > 32 sec 80 stuck requests are blocked > 4096 sec
services: mon: 3 daemons, quorum ld5505,ld5506,ld5507 (age 98m) mgr: ld5505(active, since 3d), standbys: ld5506, ld5507 mds: pve_cephfs:1 {0=ld3976=up:active} 1 up:standby osd: 368 osds: 368 up, 367 in; 302 remapped pgs
data: pools: 5 pools, 8868 pgs objects: 51.08M objects, 195 TiB usage: 590 TiB used, 563 TiB / 1.1 PiB avail pgs: 0.023% pgs not active 304136/153251211 objects degraded (0.198%) 1672190/153251211 objects misplaced (1.091%) 8564 active+clean 196 active+remapped+backfill_toofull 57 active+undersized+degraded+remapped+backfill_toofull 35 active+remapped+backfill_wait 12 active+remapped+backfill_wait+backfill_toofull 2 active+remapped+backfilling 2 peering
io: recovery: 18 MiB/s, 4 objects/s
Currently I'm using 6 OSD nodes. Node A 48x 1.6TB HDD Node B 48x 1.6TB HDD Node C 48x 1.6TB HDD Node D 48x 1.6TB HDD Node E 48x 7.2TB HDD Node F 48x 7.2TB HDD
Question: Is it advisable to distribute the drives equally over all nodes? If yes, how should this be executed w/o ceph disruption?
Regards Thomas
_______________________________________________ ceph-users mailing list -- ceph-users@ceph.io To unsubscribe send an email to ceph-users-leave@ceph.io