Hi Nico and Mark, your crush trees look indeed like they have been converted properly to using device classes already. Changing something within one device class should not influence placement in another. Maybe I'm overlooking something?
The only other place I know of where such a mix-up could occur are the crush rules. Do your rules look like this: { "rule_id": 5, "rule_name": "sr-rbd-data-one", "ruleset": 5, "type": 3, "min_size": 3, "max_size": 8, "steps": [ { "op": "set_chooseleaf_tries", "num": 50 }, { "op": "set_choose_tries", "num": 1000 }, { "op": "take", "item": -185, "item_name": "ServerRoom~rbd_data" }, { "op": "chooseleaf_indep", "num": 0, "type": "host" }, { "op": "emit" } ] } Notice the "~rbd_data" qualifier. It is important that the device class is specified at the root selection. I'm really surprised that with your crush tree you observe changes in SSD implying changes in HDD placements. I was really rough on our mimic cluster with moving disks in and out and between servers and I have never seen this problem. Could it be a regression in nautilus? Is the auto-balancer interfering? > we recently also noticed that rebuilding one pool ("ssd") > influenced speed on other pools, which was unexpected. Could this be something else? Was PG/object placement influenced or performance only? I'm asking, because during one of our service windows we observed something very strange. We have a multi-location cluster with pools with completely isolated storage devices in different locations. On one of these sub-clusters we run a ceph fs. During maintenance we needed to shut down the ceph-fs. When our admin issued the umount command (ca. 1500 clients), we noticed that RBD pools seemed to have problems even though there is absolutely no overlap in disks (disjoint crush trees), they are not even in the same physical location and sit on their own switches. The fs and RBD only share the MONs/MGRs. I'm not entirely sure if we observed something real or only a network blip. However, nagios went crazy on our VM environment for a few minutes. Maybe there is another issue that causes unexpected cross-dependencies that affect performance? Best regards, ================= Frank Schilder AIT Risø Campus Bygning 109, rum S14 ________________________________________ From: Marc Roos <m.r...@f1-outsourcing.eu> Sent: 30 September 2020 14:59:50 To: eblock; Frank Schilder Cc: ceph-users; nico.schottelius Subject: RE: [ceph-users] Re: hdd pg's migrating when converting ssd class osd's Hi Frank, thanks this 'root default' indeed looks different with these 0 there. I have also uploaded mine[1] because it looks very similar to Nico's. I guess his hdd pg's can also start moving in some occassions. Thanks for 'crushtool reclassify' hint, I guess I have missed this in the release notes or so. [1] https://pastebin.com/PFx0V3S7 -----Original Message----- To: Eugen Block Cc: Marc Roos; ceph-users Subject: Re: [ceph-users] Re: hdd pg's migrating when converting ssd class osd's This is how my crush tree including shadow hierarchies looks like (a mess :): https://pastebin.com/iCLbi4Up Every device class has its own tree. Starting with mimic, this is automatic when creating new device classes. Best regards, ================= Frank Schilder AIT Risø Campus Bygning 109, rum S14 ________________________________________ From: Eugen Block <ebl...@nde.ag> Sent: 30 September 2020 08:43:47 To: Frank Schilder Cc: Marc Roos; ceph-users Subject: Re: [ceph-users] Re: hdd pg's migrating when converting ssd class osd's Interesting, I also did this test on an upgraded cluster (L to N). I'll repeat the test on a native Nautilus to see it for myself. Zitat von Frank Schilder > Somebody on this list posted a script that can convert pre-mimic crush > trees with buckets for different types of devices to crush trees with > device classes with minimal data movement (trying to maintain IDs as > much as possible). Don't have a thread name right now, but could try > to find it tomorrow. > > I can check tomorrow how our crush tree unfolds. Basically, for every > device class there is a full copy (shadow hierarchy) for each device > class with its own weights etc. > > Best regards, > ================= > Frank Schilder > AIT Risø Campus > Bygning 109, rum S14 > > ________________________________________ > From: Marc Roos > Sent: 29 September 2020 22:19:33 > To: eblock; Frank Schilder > Cc: ceph-users > Subject: RE: [ceph-users] Re: hdd pg's migrating when converting ssd > class osd's > > Yes correct this is coming from Luminous or maybe even Kraken. How > does a default crush tree look like in mimic or octopus? Or is there > some manual how to bring this to the new 'default'? > > > -----Original Message----- > Cc: ceph-users > Subject: Re: [ceph-users] Re: hdd pg's migrating when converting ssd > class osd's > > Are these crush maps inherited from pre-mimic versions? I have > re-balanced SSD and HDD pools in mimic (mimic deployed) where one > device class never influenced the placement of the other. I have mixed > hosts and went as far as introducing rbd_meta, rbd_data and such > classes to sub-divide even further (all these devices have different perf specs). > This worked like a charm. When adding devices of one class, only pools > in this class were ever affected. > > As far as I understand, starting with mimic, every shadow class > defines a separate tree (not just leafs/OSDs). Thus, device classes > are independent of each other. > > > > ________________________________________ > Sent: 29 September 2020 20:54:48 > To: eblock > Cc: ceph-users > Subject: [ceph-users] Re: hdd pg's migrating when converting ssd class > osd's > > Yes correct, hosts have indeed both ssd's and hdd's combined. Is this > not more of a bug then? I would assume the goal of using device > classes is that you separate these and one does not affect the other, > even the host weight of the ssd and hdd class are already available. > The algorithm should just use that instead of the weight of the whole host. > Or is there some specific use case, where these classes combined is > required? > > > -----Original Message----- > Cc: ceph-users > Subject: *****SPAM***** Re: [ceph-users] Re: hdd pg's migrating when > converting ssd class osd's > > They're still in the same root (default) and each host is member of > both device-classes, I guess you have a mixed setup (hosts c01/c02 > have both HDDs and SSDs)? I don't think this separation is enough to > avoid remapping even if a different device-class is affected (your > report confirms that). > > Dividing the crush tree into different subtrees might help here but > I'm not sure if that's really something you need. You might also just > deal with the remapping as long as it doesn't happen too often, I > guess. On the other hand, if your setup won't change (except adding > more OSDs) you might as well think about a different crush tree. It > really depends on your actual requirements. > > We created two different subtrees when we got new hardware and it > helped us a lot moving the data only once to the new hardware avoiding > multiple remappings, now the older hardware is our EC environment > except for some SSDs on those old hosts that had to stay in the main > subtree. So our setup is also very individual but it works quite nice. > :-) > > > Zitat von : > >> I have practically a default setup. If I do a 'ceph osd crush tree >> --show-shadow' I have a listing like this[1]. I would assume from the >> hosts being listed within the default~ssd and default~hdd, they are >> separate (enough)? >> >> >> [1] >> root default~ssd >> host c01~ssd >> .. >> .. >> host c02~ssd >> .. >> root default~hdd >> host c01~hdd >> .. >> host c02~hdd >> .. >> root default >> >> >> >> >> -----Original Message----- >> To: ceph-users@ceph.io >> Subject: [ceph-users] Re: hdd pg's migrating when converting ssd >> class > >> osd's >> >> Are all the OSDs in the same crush root? I would think that since the >> crush weight of hosts change as soon as OSDs are out it impacts the >> whole crush tree. If you separate the SSDs from the HDDs logically > (e.g. >> different bucket type in the crush tree) the ramapping wouldn't >> affect > >> the HDDs. >> >> >> >> >>> I have been converting ssd's osd's to dmcrypt, and I have noticed >>> that >> >>> pg's of pools are migrated that should be (and are?) on hdd class. >>> >>> On a healthy ok cluster I am getting, when I set the crush reweight >>> to >> >>> 0.0 of a ssd osd this: >>> >>> 17.35 10415 0 0 9907 0 >>> 36001743890 0 0 3045 3045 >>> active+remapped+backfilling 2020-09-27 12:55:49.093054 >>> active+remapped+83758'20725398 >>> 83758:100379720 [8,14,23] 8 [3,14,23] 3 >>> 83636'20718129 2020-09-27 00:58:07.098096 83300'20689151 2020-09-24 >>> 21:42:07.385360 0 >>> >>> However osds 3,14,23,8 are all hdd osd's >>> >>> Since this is a cluster from Kraken/Luminous, I am not sure if the >>> device class of the replicated_ruleset[1] was set when the pool 17 >>> was >> >>> created. >>> Weird thing is that all pg's of this pool seem to be on hdd osd[2] >>> >>> Q. How can I display the definition of 'crush_rule 0' at the time of >>> the pool creation? (To be sure it had already this device class hdd >>> configured) >>> >>> >>> >>> [1] >>> [@~]# ceph osd pool ls detail | grep 'pool 17' >>> pool 17 'rbd' replicated size 3 min_size 2 crush_rule 0 object_hash >>> rjenkins pg_num 64 pgp_num 64 autoscale_mode warn last_change 83712 >>> flags hashpspool,selfmanaged_snaps stripe_width 0 application rbd >>> >>> >>> [@~]# ceph osd crush rule dump replicated_ruleset { >>> "rule_id": 0, >>> "rule_name": "replicated_ruleset", >>> "ruleset": 0, >>> "type": 1, >>> "min_size": 1, >>> "max_size": 10, >>> "steps": [ >>> { >>> "op": "take", >>> "item": -10, >>> "item_name": "default~hdd" >>> }, >>> { >>> "op": "chooseleaf_firstn", >>> "num": 0, >>> "type": "host" >>> }, >>> { >>> "op": "emit" >>> } >>> ] >>> } >>> >>> [2] >>> [@~]# for osd in `ceph pg dump pgs| grep '^17' | awk '{print $17" >> "$19}' >>> | grep -oE '[0-9]{1,2}'| sort -u -n`; do ceph osd crush >>> | get-device-class >>> osd.$osd ; done | sort -u >>> dumped pgs >>> hdd > > > > _______________________________________________ > ceph-users mailing list -- ceph-users@ceph.io To unsubscribe send an > email to ceph-users-le...@ceph.io _______________________________________________ ceph-users mailing list -- ceph-users@ceph.io To unsubscribe send an email to ceph-users-le...@ceph.io