Hi, Still trying to understand what is really happening under the hood, Did more test and collected the data
I Changed the `osd max pg per osd hard ratio` to 16384 but this didn't change anything
Scenario : 4 nodes , 4 disk per node , ceph 12.2.7
1. create 4 OSD with device class <pool1>
2. create pool with crush rule mapping to device class <pool1>
3. Everything OK
4. remove the 4 OSD ( delete them ) -> 0 OSD present in the system , all PG are "stalled"
5. recreate 4 new OSD ( using same disk or other disks , this doesn't matter) with ceph device class <pool2>
6. try to create pool : pool2, mapping to the device class <pool2>
Pool2 creation failure : ERANGE: pg_num 256 size 2 would mean 1024 total pgs, which exceeds max 800 (mon_max_pg_per_osd 200 * num_in_osds 4)
How did the pool check come up with 1024 total PGs when we only allocate for the pool 256 pg with size 2 ( 256*2 = 512 )
It looks like the pool check algo is calculating the total number of PG for all pools i.e. pool1 (256*2) + pool2 (256*2) rather than using the number of PG for the pool and associated OSD only
PG dump : https://paste.ee/p/sP8xZ
Ceph status :
Ceph status :
cluster:
id: ea0df043-7b25-4447-a43d-e9b2af8fe069
health: HEALTH_WARN
Reduced data availability: 37 pgs inactive, 37 pgs peering, 256 pgs stale
services:
mon: 3 daemons, quorum stratonode0.node.strato,stratonode1.node.strato,stratonode3.node.strato
mgr: stratonode0(active), standbys: stratonode1, stratonode3
osd: 4 osds: 4 up, 4 in
data:
pools: 1 pools, 256 pgs
objects: 0 objects, 0 bytes
usage: 4117 MB used, 9311 GB / 9315 GB avail
pgs: 14.453% pgs not active
219 stale+active+clean
37 stale+peering
Is there a simple way to go around this limits in the meantime? I could max out the max_pg value while retaining the 200 limit per OSD for my own PG allocation calculation
Note that allocating 2 pools, 256 PG each and replica 2 , each with their own dedicated device class mapping to theirs own 4 OSD works without any issue ( basically the above scenario without the deletion step).
It seems that it is only in the case of deleting the OSD that the whole calculation get screwed.
On Thu, 26 Jul 2018 at 20:52, Benoit Hudzia <benoit@xxxxxxxxxxxxxxx> wrote:
Sorry missing the pg dump :2.1 0 0 0 0 0 0 0 0 stale+peering 2018-07-26 19:38:13.381673 0'0 125:9 [3] 3 [3] 3 0'0 2018-07-26 15:20:08.965357 0'0 2018-07-26 15:20:08.965357 02.0 0 0 0 0 0 0 0 0 stale+peering 2018-07-26 19:38:13.345341 0'0 125:13 [3] 3 [3] 3 0'0 2018-07-26 15:20:08.965357 0'0 2018-07-26 15:20:08.965357 02 0 0 0 0 0 0 0 0sum 0 0 0 0 0 0 0 0OSD_STAT USED AVAIL TOTAL HB_PEERS PG_SUM PRIMARY_PG_SUM3 1051M 1861G 1863G [0,1,2] 256 2562 1051M 1861G 1863G [0,1,3] 0 01 1051M 3724G 3726G [0,2,3] 0 00 1051M 1861G 1863G [1,2,3] 0 0sum 4205M 9310G 9315GFor some reason it seems that some PG are allocated to osd 3 ( but stall + peering)This is kind of oddOn Thu, 26 Jul 2018 at 20:50, Benoit Hudzia <benoit@xxxxxxxxxxxxxxx> wrote:You are correct the PG are stale ( not allocated )[root@stratonode1 /]# ceph statuscluster:id: ea0df043-7b25-4447-a43d-e9b2af8fe069health: HEALTH_WARNReduced data availability: 256 pgs inactive, 256 pgs peering, 256 pgs staleservices:mon: 3 daemons, quorum stratonode1.node.strato,stratonode2.node.strato,stratonode0.node.stratomgr: stratonode1(active), standbys: stratonode2, stratonode3osd: 4 osds: 4 up, 4 indata:pools: 1 pools, 256 pgsobjects: 0 objects, 0 bytesusage: 4192 MB used, 9310 GB / 9315 GB availpgs: 100.000% pgs not active256 stale+peeringPG dump : show all PG in stale + peeringHowever it s kind of strange it show some PG associated with OSD 3So it seems that PGcalc is not taking into account the ruleset .....Do you think that changing ""osd max pg per osd hard ratio"" to a huge number (1M) would be a valid temp workaround ?We always allocate pool with dedicated OSD using the device class rule set , so we never have pool sharing OSD .
I ll open a bug with ceph regarding pg creation check ignoring the crush ruleset.On Thu, 26 Jul 2018 at 17:11, John Spray <jspray@xxxxxxxxxx> wrote:On Thu, Jul 26, 2018 at 4:57 PM Benoit Hudzia <benoit@xxxxxxxxxxxxxxx> wrote:HI,We currently segregate ceph pool PG allocation using the crush device class ruleset as described: https://ceph.com/community/new-luminous-crush-device-classes/simply using the following command to define the rule : ceph osd crush rule create-replicated <RULE> default host <DEVICE CLASS>However, we noticed that the rule is not strict in certain scenarios. By that, I mean that if there is no OSD of the specific device class ceph will allocate PG for this pool to any other OSD available ( creating an issue with the PG calculation when we want to add new pool)Simple scenario :1. create 1 Pool : <pool1> , replication 2 with 4 nodes , 1 OSD each . belonging to class <pool1>2. remove all OSD ( delete them )3. create 4 new OSD (using same disk but different ID) but this time tag them with class <pool2>4. Try to create pool <pool2> -> this will fail withthe pool creation will fail with output : Error ERANGE: pg_num 256 size 2 would mean 1024 total pgs, which exceeds max 800 (mon_max_pg_per_osd 200 * num_in_osds 4)"Pool1 simply started allocating PG to OSD that doesn't belong to the rulesetAre you sure pool 1's PGs are actually being placed on the wrong OSDs? Have you looked at the output of "ceph pg dump" to check that?It sounds more like the pool creation check is simply failing to consider the crush rules and applying a cruder global check.John_______________________________________________Which leads me to the following question: is there a way to make the crush rule a hard requirement. E.g : if we do not have any osd matching the device class , it won't start trying to allocate pg to OSD that doesn't match it?Is there any way to prevent pool 1 to use the OSD ?
ceph-users mailing list
ceph-users@xxxxxxxxxxxxxx
http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
----
_______________________________________________ ceph-users mailing list ceph-users@xxxxxxxxxxxxxx http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
