Hi, I am just evaluating out cluster configuration again, because we had an very bad incident with laggy OSDs that shut down the entire cluster.
We use datacenter SSDs in different sizes (2, 4, 8TB) and someone said, that I should not go beyond a specific amount of PGs on certain device classes. Now the distribution in our cluster is 100PGs on the 2TB disks, 200PGs on the 4TB disks and 400PGs on the 8TB disks. According to this ( https://docs.ceph.com/en/latest/rados/operations/placement-groups/#choosing-the-number-of-placement-groups) I should got with 150OSDs * 100 / 3 replication and round UP to the nearest power of two which would be 8k BUT because these are just SSDs (100k/30k read/write IOPS) I think I should got down to 4k PGs, to not overload the 8TB disks with operations. What do you think? These are out cluster stats: # ceph df --- RAW STORAGE --- CLASS SIZE AVAIL USED RAW USED %RAW USED ssd 419 TiB 122 TiB 297 TiB 298 TiB 71.01 TOTAL 419 TiB 122 TiB 297 TiB 298 TiB 71.01 --- POOLS --- POOL ID PGS STORED OBJECTS USED %USED MAX AVAIL isos 7 64 431 GiB 111.58k 1.3 TiB 1.57 26 TiB rbd 8 8192 98 TiB 34.17M 288 TiB 78.46 26 TiB archive 9 128 2.6 TiB 722.10k 7.8 TiB 8.99 26 TiB device_health_metrics 10 1 203 MiB 156 608 MiB 0 26 TiB # ceph status cluster: id: 74313356-3b3d-43f3-bce6-9fb0e4591097 health: HEALTH_OK services: mon: 3 daemons, quorum ceph-rbd-mon4,ceph-rbd-mon5,ceph-rbd-mon6 (age 44h) mgr: ceph-rbd-mon5(active, since 44h), standbys: ceph-rbd-mon6, ceph-rbd-mon4 osd: 150 osds: 150 up (since 44h), 150 in (since 6d) data: pools: 4 pools, 8385 pgs objects: 35.01M objects, 110 TiB usage: 298 TiB used, 122 TiB / 419 TiB avail pgs: 8380 active+clean 3 active+clean+scrubbing+deep 2 active+clean+scrubbing io: client: 100 MiB/s rd, 433 MiB/s wr, 1.45k op/s rd, 12.02k op/s wr -- Die Selbsthilfegruppe "UTF-8-Probleme" trifft sich diesmal abweichend im groüen Saal. _______________________________________________ ceph-users mailing list -- ceph-users@ceph.io To unsubscribe send an email to ceph-users-le...@ceph.io
