[ceph-users] Re: cephadm automatic sizing of WAL/DB on SSD
I think that the backported fix for this issue made it into ceph v16.2.11. https://ceph.io/en/news/blog/2023/v16-2-11-pacific-released/ "ceph-volume: Pacific backports (pr#47413, Guillaume Abrioux, Zack Cerza, Arthur Outhenin-Chalandre)" https://github.com/ceph/ceph/pull/47413/commits/4252cc44211f0ccebf388374744eaa26b32854d3 -Patrick ___ ceph-users mailing list -- ceph-users@ceph.io To unsubscribe send an email to ceph-users-le...@ceph.io
[ceph-users] Re: cephadm automatic sizing of WAL/DB on SSD
Hi, We were also affected by this bug when we deployed a new Pacific cluster. Any news about the release of this fix to Ceph Pacific? It looks done for Quincy version but not Pacific. https://github.com/ceph/ceph/pull/47292 Regards, Adrien Le 05/10/2022 à 13:21, Anh Phan Tuan a écrit : It seems the 17.2.4 release has fixed this. ceph-volume: fix fast device alloc size on mulitple device (pr#47293, Arthur Outhenin-Chalandre) Bug #56031: batch compute a lower size than what it should be for blockdb with multiple fast device - ceph-volume - Ceph <https://tracker.ceph.com/issues/56031> Regards, Anh Phan On Fri, Sep 16, 2022 at 2:34 AM Christophe BAILLON wrote: Hi The problem is still present in version 17.2.3, thanks for the trick to work around... Regards - Mail original - De: "Anh Phan Tuan" À: "Calhoun, Patrick" Cc: "Arthur Outhenin-Chalandre" , "ceph-users" Envoyé: Jeudi 11 Août 2022 10:14:17 Objet: [ceph-users] Re: cephadm automatic sizing of WAL/DB on SSD Hi Patrick, I am also facing this bug when deploying a new cluster at the time 16.2.7 release. The bugs relative to the way ceph calculator db_size form give db disk. Instead of : slot db size = size of db disk / num slot per disk. Ceph calculated the value: slot db size = size of db disk (just one disk) / total number of slots needed (number of osd prepared in that time). In your case, you have 2 db disks, It will make the db size only 50% of the corrected value. In my case, I have 4 db disks per host, It makes the db size only 25% of the corrected value. This bug happens even when you deploy by batch command. In that time, I finally used to work around by batch command but only deploy all osd relative to one db disk a time, in this case ceph calculated the correct value. Cheers, Anh Phan On Sat, Jul 30, 2022 at 12:31 AM Calhoun, Patrick wrote: Thanks, Arthur, I think you are right about that bug looking very similar to what I've observed. I'll try to remember to update the list once the fix is merged and released and I get a chance to test it. I'm hoping somebody can comment on what are ceph's current best practices for sizing WAL/DB volumes, considering rocksdb levels and compaction. -Patrick From: Arthur Outhenin-Chalandre Sent: Friday, July 29, 2022 2:11 AM To: ceph-users@ceph.io Subject: [ceph-users] Re: cephadm automatic sizing of WAL/DB on SSD Hi Patrick, On 7/28/22 16:22, Calhoun, Patrick wrote: In a new OSD node with 24 hdd (16 TB each) and 2 ssd (1.44 TB each), I'd like to have "ceph orch" allocate WAL and DB on the ssd devices. I use the following service spec: spec: data_devices: rotational: 1 size: '14T:' db_devices: rotational: 0 size: '1T:' db_slots: 12 This results in each OSD having a 60GB volume for WAL/DB, which equates to 50% total usage in the VG on each ssd, and 50% free. I honestly don't know what size to expect, but exactly 50% of capacity makes me suspect this is due to a bug: https://tracker.ceph.com/issues/54541 (In fact, I had run into this bug when specifying block_db_size rather than db_slots) Questions: Am I being bit by that bug? Is there a better approach, in general, to my situation? Are DB sizes still governed by the rocksdb tiering? (I thought that this was mostly resolved by https://github.com/ceph/ceph/pull/29687 ) If I provision a DB/WAL logical volume size to 61GB, is that effectively a 30GB database, and 30GB of extra room for compaction? I don't use cephadm, but it's maybe related to this regression: https://tracker.ceph.com/issues/56031. At list the symptoms looks very similar... Cheers, -- Arthur Outhenin-Chalandre ___ 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 ___ ceph-users mailing list -- ceph-users@ceph.io To unsubscribe send an email to ceph-users-le...@ceph.io -- Christophe BAILLON Mobile :: +336 16 400 522 Work :: https://eyona.com Twitter :: https://twitter.com/ctof ___ 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
[ceph-users] Re: cephadm automatic sizing of WAL/DB on SSD
It seems the 17.2.4 release has fixed this. ceph-volume: fix fast device alloc size on mulitple device (pr#47293, > Arthur Outhenin-Chalandre) Bug #56031: batch compute a lower size than what it should be for blockdb with multiple fast device - ceph-volume - Ceph <https://tracker.ceph.com/issues/56031> Regards, Anh Phan On Fri, Sep 16, 2022 at 2:34 AM Christophe BAILLON wrote: > Hi > > The problem is still present in version 17.2.3, > thanks for the trick to work around... > > Regards > > - Mail original - > > De: "Anh Phan Tuan" > > À: "Calhoun, Patrick" > > Cc: "Arthur Outhenin-Chalandre" , > "ceph-users" > > Envoyé: Jeudi 11 Août 2022 10:14:17 > > Objet: [ceph-users] Re: cephadm automatic sizing of WAL/DB on SSD > > > Hi Patrick, > > > > I am also facing this bug when deploying a new cluster at the time 16.2.7 > > release. > > > > The bugs relative to the way ceph calculator db_size form give db disk. > > > > Instead of : slot db size = size of db disk / num slot per disk. > > Ceph calculated the value: slot db size = size of db disk (just one > disk) / > > total number of slots needed (number of osd prepared in that time). > > > > In your case, you have 2 db disks, It will make the db size only 50% of > the > > corrected value. > > In my case, I have 4 db disks per host, It makes the db size only 25% of > > the corrected value. > > > > This bug happens even when you deploy by batch command. > > In that time, I finally used to work around by batch command but only > > deploy all osd relative to one db disk a time, in this case ceph > calculated > > the correct value. > > > > Cheers, > > Anh Phan > > > > > > > > On Sat, Jul 30, 2022 at 12:31 AM Calhoun, Patrick > wrote: > > > >> Thanks, Arthur, > >> > >> I think you are right about that bug looking very similar to what I've > >> observed. I'll try to remember to update the list once the fix is merged > >> and released and I get a chance to test it. > >> > >> I'm hoping somebody can comment on what are ceph's current best > practices > >> for sizing WAL/DB volumes, considering rocksdb levels and compaction. > >> > >> -Patrick > >> > >> > >> From: Arthur Outhenin-Chalandre > >> Sent: Friday, July 29, 2022 2:11 AM > >> To: ceph-users@ceph.io > >> Subject: [ceph-users] Re: cephadm automatic sizing of WAL/DB on SSD > >> > >> Hi Patrick, > >> > >> On 7/28/22 16:22, Calhoun, Patrick wrote: > >> > In a new OSD node with 24 hdd (16 TB each) and 2 ssd (1.44 TB each), > I'd > >> like to have "ceph orch" allocate WAL and DB on the ssd devices. > >> > > >> > I use the following service spec: > >> > spec: > >> > data_devices: > >> > rotational: 1 > >> > size: '14T:' > >> > db_devices: > >> > rotational: 0 > >> > size: '1T:' > >> > db_slots: 12 > >> > > >> > This results in each OSD having a 60GB volume for WAL/DB, which > equates > >> to 50% total usage in the VG on each ssd, and 50% free. > >> > I honestly don't know what size to expect, but exactly 50% of capacity > >> makes me suspect this is due to a bug: > >> > https://tracker.ceph.com/issues/54541 > >> > (In fact, I had run into this bug when specifying block_db_size rather > >> than db_slots) > >> > > >> > Questions: > >> > Am I being bit by that bug? > >> > Is there a better approach, in general, to my situation? > >> > Are DB sizes still governed by the rocksdb tiering? (I thought that > >> this was mostly resolved by https://github.com/ceph/ceph/pull/29687 ) > >> > If I provision a DB/WAL logical volume size to 61GB, is that > >> effectively a 30GB database, and 30GB of extra room for compaction? > >> > >> I don't use cephadm, but it's maybe related to this regression: > >> https://tracker.ceph.com/issues/56031. At list the symptoms looks very > >> similar... > >> > >> Cheers, > >> > >> -- > >> Arthur Outhenin-Chalandre > >> ___ > >> 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 > >> > > ___ > > ceph-users mailing list -- ceph-users@ceph.io > > To unsubscribe send an email to ceph-users-le...@ceph.io > > -- > Christophe BAILLON > Mobile :: +336 16 400 522 > Work :: https://eyona.com > Twitter :: https://twitter.com/ctof > ___ ceph-users mailing list -- ceph-users@ceph.io To unsubscribe send an email to ceph-users-le...@ceph.io
[ceph-users] Re: cephadm automatic sizing of WAL/DB on SSD
Hi The problem is still present in version 17.2.3, thanks for the trick to work around... Regards - Mail original - > De: "Anh Phan Tuan" > À: "Calhoun, Patrick" > Cc: "Arthur Outhenin-Chalandre" , > "ceph-users" > Envoyé: Jeudi 11 Août 2022 10:14:17 > Objet: [ceph-users] Re: cephadm automatic sizing of WAL/DB on SSD > Hi Patrick, > > I am also facing this bug when deploying a new cluster at the time 16.2.7 > release. > > The bugs relative to the way ceph calculator db_size form give db disk. > > Instead of : slot db size = size of db disk / num slot per disk. > Ceph calculated the value: slot db size = size of db disk (just one disk) / > total number of slots needed (number of osd prepared in that time). > > In your case, you have 2 db disks, It will make the db size only 50% of the > corrected value. > In my case, I have 4 db disks per host, It makes the db size only 25% of > the corrected value. > > This bug happens even when you deploy by batch command. > In that time, I finally used to work around by batch command but only > deploy all osd relative to one db disk a time, in this case ceph calculated > the correct value. > > Cheers, > Anh Phan > > > > On Sat, Jul 30, 2022 at 12:31 AM Calhoun, Patrick wrote: > >> Thanks, Arthur, >> >> I think you are right about that bug looking very similar to what I've >> observed. I'll try to remember to update the list once the fix is merged >> and released and I get a chance to test it. >> >> I'm hoping somebody can comment on what are ceph's current best practices >> for sizing WAL/DB volumes, considering rocksdb levels and compaction. >> >> -Patrick >> >> >> From: Arthur Outhenin-Chalandre >> Sent: Friday, July 29, 2022 2:11 AM >> To: ceph-users@ceph.io >> Subject: [ceph-users] Re: cephadm automatic sizing of WAL/DB on SSD >> >> Hi Patrick, >> >> On 7/28/22 16:22, Calhoun, Patrick wrote: >> > In a new OSD node with 24 hdd (16 TB each) and 2 ssd (1.44 TB each), I'd >> like to have "ceph orch" allocate WAL and DB on the ssd devices. >> > >> > I use the following service spec: >> > spec: >> > data_devices: >> > rotational: 1 >> > size: '14T:' >> > db_devices: >> > rotational: 0 >> > size: '1T:' >> > db_slots: 12 >> > >> > This results in each OSD having a 60GB volume for WAL/DB, which equates >> to 50% total usage in the VG on each ssd, and 50% free. >> > I honestly don't know what size to expect, but exactly 50% of capacity >> makes me suspect this is due to a bug: >> > https://tracker.ceph.com/issues/54541 >> > (In fact, I had run into this bug when specifying block_db_size rather >> than db_slots) >> > >> > Questions: >> > Am I being bit by that bug? >> > Is there a better approach, in general, to my situation? >> > Are DB sizes still governed by the rocksdb tiering? (I thought that >> this was mostly resolved by https://github.com/ceph/ceph/pull/29687 ) >> > If I provision a DB/WAL logical volume size to 61GB, is that >> effectively a 30GB database, and 30GB of extra room for compaction? >> >> I don't use cephadm, but it's maybe related to this regression: >> https://tracker.ceph.com/issues/56031. At list the symptoms looks very >> similar... >> >> Cheers, >> >> -- >> Arthur Outhenin-Chalandre >> ___ >> 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 >> > ___ > ceph-users mailing list -- ceph-users@ceph.io > To unsubscribe send an email to ceph-users-le...@ceph.io -- Christophe BAILLON Mobile :: +336 16 400 522 Work :: https://eyona.com Twitter :: https://twitter.com/ctof ___ ceph-users mailing list -- ceph-users@ceph.io To unsubscribe send an email to ceph-users-le...@ceph.io
[ceph-users] Re: cephadm automatic sizing of WAL/DB on SSD
Hi Patrick, I am also facing this bug when deploying a new cluster at the time 16.2.7 release. The bugs relative to the way ceph calculator db_size form give db disk. Instead of : slot db size = size of db disk / num slot per disk. Ceph calculated the value: slot db size = size of db disk (just one disk) / total number of slots needed (number of osd prepared in that time). In your case, you have 2 db disks, It will make the db size only 50% of the corrected value. In my case, I have 4 db disks per host, It makes the db size only 25% of the corrected value. This bug happens even when you deploy by batch command. In that time, I finally used to work around by batch command but only deploy all osd relative to one db disk a time, in this case ceph calculated the correct value. Cheers, Anh Phan On Sat, Jul 30, 2022 at 12:31 AM Calhoun, Patrick wrote: > Thanks, Arthur, > > I think you are right about that bug looking very similar to what I've > observed. I'll try to remember to update the list once the fix is merged > and released and I get a chance to test it. > > I'm hoping somebody can comment on what are ceph's current best practices > for sizing WAL/DB volumes, considering rocksdb levels and compaction. > > -Patrick > > > From: Arthur Outhenin-Chalandre > Sent: Friday, July 29, 2022 2:11 AM > To: ceph-users@ceph.io > Subject: [ceph-users] Re: cephadm automatic sizing of WAL/DB on SSD > > Hi Patrick, > > On 7/28/22 16:22, Calhoun, Patrick wrote: > > In a new OSD node with 24 hdd (16 TB each) and 2 ssd (1.44 TB each), I'd > like to have "ceph orch" allocate WAL and DB on the ssd devices. > > > > I use the following service spec: > > spec: > > data_devices: > > rotational: 1 > > size: '14T:' > > db_devices: > > rotational: 0 > > size: '1T:' > > db_slots: 12 > > > > This results in each OSD having a 60GB volume for WAL/DB, which equates > to 50% total usage in the VG on each ssd, and 50% free. > > I honestly don't know what size to expect, but exactly 50% of capacity > makes me suspect this is due to a bug: > > https://tracker.ceph.com/issues/54541 > > (In fact, I had run into this bug when specifying block_db_size rather > than db_slots) > > > > Questions: > > Am I being bit by that bug? > > Is there a better approach, in general, to my situation? > > Are DB sizes still governed by the rocksdb tiering? (I thought that > this was mostly resolved by https://github.com/ceph/ceph/pull/29687 ) > > If I provision a DB/WAL logical volume size to 61GB, is that > effectively a 30GB database, and 30GB of extra room for compaction? > > I don't use cephadm, but it's maybe related to this regression: > https://tracker.ceph.com/issues/56031. At list the symptoms looks very > similar... > > Cheers, > > -- > Arthur Outhenin-Chalandre > ___ > 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 > ___ ceph-users mailing list -- ceph-users@ceph.io To unsubscribe send an email to ceph-users-le...@ceph.io
[ceph-users] Re: cephadm automatic sizing of WAL/DB on SSD
Thanks, Arthur, I think you are right about that bug looking very similar to what I've observed. I'll try to remember to update the list once the fix is merged and released and I get a chance to test it. I'm hoping somebody can comment on what are ceph's current best practices for sizing WAL/DB volumes, considering rocksdb levels and compaction. -Patrick From: Arthur Outhenin-Chalandre Sent: Friday, July 29, 2022 2:11 AM To: ceph-users@ceph.io Subject: [ceph-users] Re: cephadm automatic sizing of WAL/DB on SSD Hi Patrick, On 7/28/22 16:22, Calhoun, Patrick wrote: > In a new OSD node with 24 hdd (16 TB each) and 2 ssd (1.44 TB each), I'd like > to have "ceph orch" allocate WAL and DB on the ssd devices. > > I use the following service spec: > spec: > data_devices: > rotational: 1 > size: '14T:' > db_devices: > rotational: 0 > size: '1T:' > db_slots: 12 > > This results in each OSD having a 60GB volume for WAL/DB, which equates to > 50% total usage in the VG on each ssd, and 50% free. > I honestly don't know what size to expect, but exactly 50% of capacity makes > me suspect this is due to a bug: > https://tracker.ceph.com/issues/54541 > (In fact, I had run into this bug when specifying block_db_size rather than > db_slots) > > Questions: > Am I being bit by that bug? > Is there a better approach, in general, to my situation? > Are DB sizes still governed by the rocksdb tiering? (I thought that this > was mostly resolved by https://github.com/ceph/ceph/pull/29687 ) > If I provision a DB/WAL logical volume size to 61GB, is that effectively a > 30GB database, and 30GB of extra room for compaction? I don't use cephadm, but it's maybe related to this regression: https://tracker.ceph.com/issues/56031. At list the symptoms looks very similar... Cheers, -- Arthur Outhenin-Chalandre ___ 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
[ceph-users] Re: cephadm automatic sizing of WAL/DB on SSD
Hi Patrick, On 7/28/22 16:22, Calhoun, Patrick wrote: > In a new OSD node with 24 hdd (16 TB each) and 2 ssd (1.44 TB each), I'd like > to have "ceph orch" allocate WAL and DB on the ssd devices. > > I use the following service spec: > spec: > data_devices: > rotational: 1 > size: '14T:' > db_devices: > rotational: 0 > size: '1T:' > db_slots: 12 > > This results in each OSD having a 60GB volume for WAL/DB, which equates to > 50% total usage in the VG on each ssd, and 50% free. > I honestly don't know what size to expect, but exactly 50% of capacity makes > me suspect this is due to a bug: > https://tracker.ceph.com/issues/54541 > (In fact, I had run into this bug when specifying block_db_size rather than > db_slots) > > Questions: > Am I being bit by that bug? > Is there a better approach, in general, to my situation? > Are DB sizes still governed by the rocksdb tiering? (I thought that this > was mostly resolved by https://github.com/ceph/ceph/pull/29687 ) > If I provision a DB/WAL logical volume size to 61GB, is that effectively a > 30GB database, and 30GB of extra room for compaction? I don't use cephadm, but it's maybe related to this regression: https://tracker.ceph.com/issues/56031. At list the symptoms looks very similar... Cheers, -- Arthur Outhenin-Chalandre ___ ceph-users mailing list -- ceph-users@ceph.io To unsubscribe send an email to ceph-users-le...@ceph.io