|
I previously enabled upmap and used automatic balancing with "ceph balancer on". I got very good results and OSD's ended up with perfectly distributed pg's. Now after adding several new OSD's, auto balancing does not seem to be working anymore. OSD's have 30-50% usage where previously all had almost the same %. I turned off auto balancer and tried manually running a plan # ceph balancer reset # ceph balancer optimize myplan # ceph balancer show myplan ceph osd pg-upmap-items 41.1 106 125 95 121 84 34 36 99 72 126 ceph osd pg-upmap-items 41.5 12 121 65 3 122 52 5 126 ceph osd pg-upmap-items 41.b 117 99 65 125 ceph osd pg-upmap-items 41.c 49 121 81 131 ceph osd pg-upmap-items 41.e 61 82 73 52 122 46 84 118 ceph osd pg-upmap-items 41.f 71 127 15 121 56 82 ceph osd pg-upmap-items 41.12 81 92 ceph osd pg-upmap-items 41.17 35 127 71 44 ceph osd pg-upmap-items 41.19 81 131 21 119 18 52 ceph osd pg-upmap-items 41.25 18 52 37 125 40 3 41 34 71 127 4 128 After running this plan there's no difference and still huge inbalance on the OSD's. Creating a new plan give the same plan again. # ceph balancer eval current cluster score 0.015162 (lower is better) Balancer eval shows quite low number, so it seems to think the pg distribution is already optimized ? Since i'm not getting this working again. I looked into the offline optimization at http://docs.ceph.com/docs/mimic/rados/operations/upmap/ I have 2 pools. Replicated pool using 3 OSD's with "10k" device class. And remaining OSD's have "hdd" device class. The resulting out.txt creates a much larger plan, but would map alot of PG's to the "10k" OSD's (where they should not be). And i can't seem to find any way to exclude these 3 OSD's. Any ideas how to proceed ? Please, paste (on pastebin) your `ceph osd df tree`. What is your ceph version? Also you can enable balancer debug messages if set `debug mgr = 4/5` in your ceph.conf
k |
_______________________________________________ ceph-users mailing list ceph-users@xxxxxxxxxxxxxx http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
