On Mon, Mar 25, 2024 at 7:37 PM Torkil Svensgaard <torkil@drcmr.dk> wrote:
On 24/03/2024 01:14, Torkil Svensgaard wrote:
On 24-03-2024 00:31, Alexander E. Patrakov wrote:
Hi Torkil,
Hi Alexander
Thanks for the update. Even though the improvement is small, it is still an improvement, consistent with the osd_max_backfills value, and it proves that there are still unsolved peering issues.
I have looked at both the old and the new state of the PG, but could not find anything else interesting.
I also looked again at the state of PG 37.1. It is known what blocks the backfill of this PG; please search for "blocked_by." However, this is just one data point, which is insufficient for any conclusions. Try looking at other PGs. Is there anything too common in the non-empty "blocked_by" blocks?
I'll take a look at that tomorrow, perhaps we can script something meaningful.
Hi Alexander
While working on a script querying all PGs and making a list of all OSDs found in a blocked_by list, and how many times for each, I discovered something odd about pool 38:
" [root@lazy blocked_by]# sh blocked_by.sh 38 |tee pool38 OSDs blocking other OSDs:
<snip>
All PGs in the pool are active+clean so why are there any blocked_by at all? One example attached.
I don't know. In any case, it doesn't match the "one OSD blocks them all" scenario that I was looking for. I think this is something bogus that can probably be cleared in your example by restarting osd.89 (i.e, the one being blocked).
Mvh.
Torkil
I think we have to look for patterns in other ways, too. One tool that produces good visualizations is TheJJ balancer. Although it is called a "balancer," it can also visualize the ongoing backfills.
The tool is available at https://raw.githubusercontent.com/TheJJ/ceph-balancer/master/placementoptimi...
Run it as follows:
./placementoptimizer.py showremapped --by-osd | tee remapped.txt
Output attached.
Thanks again.
Mvh.
Torkil
On Sun, Mar 24, 2024 at 5:50 AM Torkil Svensgaard <torkil@drcmr.dk> wrote:
Hi Alex
New query output attached after restarting both OSDs. OSD 237 is no longer mentioned but it unfortunately made no difference for the number of backfills which went 59->62->62.
Mvh.
Torkil
On 23-03-2024 22:26, Alexander E. Patrakov wrote:
Hi Torkil,
I have looked at the files that you attached. They were helpful: pool 11 is problematic, it complains about degraded objects for no obvious reason. I think that is the blocker.
I also noted that you mentioned peering problems, and I suspect that they are not completely resolved. As a somewhat-irrational move, to confirm this theory, you can restart osd.237 (it is mentioned at the end of query.11.fff.txt, although I don't understand why it is there) and then osd.298 (it is the primary for that pg) and see if any additional backfills are unblocked after that. Also, please re-query that PG again after the OSD restart.
On Sun, Mar 24, 2024 at 4:56 AM Torkil Svensgaard <torkil@drcmr.dk> wrote:
On 23-03-2024 21:19, Alexander E. Patrakov wrote: > Hi Torkil,
Hi Alexander
> I have looked at the CRUSH rules, and the equivalent rules work on my > test cluster. So this cannot be the cause of the blockage.
Thank you for taking the time =)
> What happens if you increase the osd_max_backfills setting > temporarily?
We already had the mclock override option in place and I re-enabled our babysitter script which sets osd_max_backfills pr OSD to 1-3 depending on how full they are. Active backfills went from 16 to 53 which is probably because default osd_max_backfills for mclock is 1.
I think 53 is still a low number of active backfills given the large percentage misplaced.
> It may be a good idea to investigate a few of the stalled PGs. Please > run commands similar to this one: > > ceph pg 37.0 query > query.37.0.txt > ceph pg 37.1 query > query.37.1.txt > ... > and the same for the other affected pools.
A few samples attached.
> Still, I must say that some of your rules are actually unsafe. > > The 4+2 rule as used by rbd_ec_data will not survive a > datacenter-offline incident. Namely, for each PG, it chooses OSDs > from > two hosts in each datacenter, so 6 OSDs total. When a datacenter is > offline, you will, therefore, have only 4 OSDs up, which is exactly > the number of data chunks. However, the pool requires min_size 5, so > all PGs will be inactive (to prevent data corruption) and will stay > inactive until the datacenter comes up again. However, please don't > set min_size to 4 - then, any additional incident (like a defective > disk) will lead to data loss, and the shards in the datacenter which > went offline would be useless because they do not correspond to the > updated shards written by the clients.
Thanks for the explanation. This is an old pool predating the 3 DC setup and we'll migrate the data to a 4+5 pool when we can.
> The 4+5 rule as used by cephfs.hdd.data has min_size equal to the > number of data chunks. See above why it is bad. Please set > min_size to > 5.
Thanks, that was a leftover for getting the PGs to peer (stuck at creating+incomplete) when we created the pool. It's back to 5 now.
> The rbd.ssd.data pool seems to be OK - and, by the way, its PGs are > 100% active+clean.
There is very little data in this pool, that is probably the main reason.
> Regarding the mon_max_pg_per_osd setting, you have a few OSDs that > have 300+ PGs, the observed maximum is 347. Please set it to 400.
Copy that. Didn't seem to make a difference though, and we have osd_max_pg_per_osd_hard_ratio set to 5.000000.
Mvh.
Torkil
> On Sun, Mar 24, 2024 at 3:16 AM Torkil Svensgaard > <torkil@drcmr.dk> wrote: >> >> >> >> On 23-03-2024 19:05, Alexander E. Patrakov wrote: >>> Sorry for replying to myself, but "ceph osd pool ls detail" by >>> itself >>> is insufficient. For every erasure code profile mentioned in the >>> output, please also run something like this: >>> >>> ceph osd erasure-code-profile get prf-for-ec-data >>> >>> ...where "prf-for-ec-data" is the name that appears after the words >>> "erasure profile" in the "ceph osd pool ls detail" output. >> >> [root@lazy ~]# ceph osd pool ls detail | grep erasure >> pool 11 'rbd_ec_data' erasure profile DRCMR_k4m2 size 6 min_size 5 >> crush_rule 0 object_hash rjenkins pg_num 4096 pgp_num 4096 >> autoscale_mode off last_change 2257933 lfor 0/1291190/1755101 flags >> hashpspool,ec_overwrites,selfmanaged_snaps,bulk stripe_width 16384 >> fast_read 1 compression_algorithm snappy compression_mode aggressive >> application rbd >> pool 37 'cephfs.hdd.data' erasure profile >> DRCMR_k4m5_datacenter_hdd size >> 9 min_size 4 crush_rule 7 object_hash rjenkins pg_num 2048 >> pgp_num 2048 >> autoscale_mode off last_change 2257933 lfor 0/0/2139486 flags >> hashpspool,ec_overwrites,bulk stripe_width 16384 fast_read 1 >> compression_algorithm zstd compression_mode aggressive >> application cephfs >> pool 38 'rbd.ssd.data' erasure profile DRCMR_k4m5_datacenter_ssd >> size 9 >> min_size 5 crush_rule 8 object_hash rjenkins pg_num 32 pgp_num 32 >> autoscale_mode warn last_change 2198930 lfor 0/2198930/2198928 flags >> hashpspool,ec_overwrites,selfmanaged_snaps stripe_width 16384 >> compression_algorithm zstd compression_mode aggressive >> application rbd >> >> [root@lazy ~]# ceph osd erasure-code-profile get DRCMR_k4m2 >> crush-device-class=hdd >> crush-failure-domain=host >> crush-root=default >> jerasure-per-chunk-alignment=false >> k=4 >> m=2 >> plugin=jerasure >> technique=reed_sol_van >> w=8 >> [root@lazy ~]# ceph osd erasure-code-profile get >> DRCMR_k4m5_datacenter_hdd >> crush-device-class=hdd >> crush-failure-domain=datacenter >> crush-root=default >> jerasure-per-chunk-alignment=false >> k=4 >> m=5 >> plugin=jerasure >> technique=reed_sol_van >> w=8 >> [root@lazy ~]# ceph osd erasure-code-profile get >> DRCMR_k4m5_datacenter_ssd >> crush-device-class=ssd >> crush-failure-domain=datacenter >> crush-root=default >> jerasure-per-chunk-alignment=false >> k=4 >> m=5 >> plugin=jerasure >> technique=reed_sol_van >> w=8 >> >> But as I understand it those profiles are only used to create the >> initial crush rule for the pool, and we have manually edited >> those along >> the way. Here are the 3 rules in use for the 3 EC pools: >> >> rule rbd_ec_data { >> id 0 >> type erasure >> step set_chooseleaf_tries 5 >> step set_choose_tries 100 >> step take default class hdd >> step choose indep 0 type datacenter >> step chooseleaf indep 2 type host >> step emit >> } >> rule cephfs.hdd.data { >> id 7 >> type erasure >> step set_chooseleaf_tries 5 >> step set_choose_tries 100 >> step take default class hdd >> step choose indep 0 type datacenter >> step chooseleaf indep 3 type host >> step emit >> } >> rule rbd.ssd.data { >> id 8 >> type erasure >> step set_chooseleaf_tries 5 >> step set_choose_tries 100 >> step take default class ssd >> step choose indep 0 type datacenter >> step chooseleaf indep 3 type host >> step emit >> } >> >> Which should first pick all 3 datacenters in the choose step and >> then >> either 2 or 3 hosts in the chooseleaf step, matching EC 4+2 and 4+5 >> respectively. >> >> Mvh. >> >> Torkil >> >>> On Sun, Mar 24, 2024 at 1:56 AM Alexander E. Patrakov >>> <patrakov@gmail.com> wrote: >>>> >>>> Hi Torkil, >>>> >>>> I take my previous response back. >>>> >>>> You have an erasure-coded pool with nine shards but only three >>>> datacenters. This, in general, cannot work. You need either nine >>>> datacenters or a very custom CRUSH rule. The second option may >>>> not be >>>> available if the current EC setup is already incompatible, as >>>> there is >>>> no way to change the EC parameters. >>>> >>>> It would help if you provided the output of "ceph osd pool ls >>>> detail". >>>> >>>> On Sun, Mar 24, 2024 at 1:43 AM Alexander E. Patrakov >>>> <patrakov@gmail.com> wrote: >>>>> >>>>> Hi Torkil, >>>>> >>>>> Unfortunately, your files contain nothing obviously bad or >>>>> suspicious, >>>>> except for two things: more PGs than usual and bad balance. >>>>> >>>>> What's your "mon max pg per osd" setting? >>>>> >>>>> On Sun, Mar 24, 2024 at 1:08 AM Torkil Svensgaard >>>>> <torkil@drcmr.dk> wrote: >>>>>> >>>>>> On 2024-03-23 17:54, Kai Stian Olstad wrote: >>>>>>> On Sat, Mar 23, 2024 at 12:09:29PM +0100, Torkil Svensgaard >>>>>>> wrote: >>>>>>>> >>>>>>>> The other output is too big for pastebin and I'm not >>>>>>>> familiar with >>>>>>>> paste services, any suggestion for a preferred way to share >>>>>>>> such >>>>>>>> output? >>>>>>> >>>>>>> You can attached files to the mail here on the list. >>>>>> >>>>>> Doh, for some reason I was sure attachments would be >>>>>> stripped. Thanks, >>>>>> attached. >>>>>> >>>>>> Mvh. >>>>>> >>>>>> Torkil >>>>> >>>>> >>>>> >>>>> -- >>>>> Alexander E. Patrakov >>>> >>>> >>>> >>>> -- >>>> Alexander E. Patrakov >>> >>> >>> >> >> -- >> Torkil Svensgaard >> Systems Administrator >> Danish Research Centre for Magnetic Resonance DRCMR, Section 714 >> Copenhagen University Hospital Amager and Hvidovre >> Kettegaard Allé 30, 2650 Hvidovre, Denmark >> > >
-- Torkil Svensgaard Systems Administrator Danish Research Centre for Magnetic Resonance DRCMR, Section 714 Copenhagen University Hospital Amager and Hvidovre Kettegaard Allé 30, 2650 Hvidovre, Denmark
-- Torkil Svensgaard Systems Administrator Danish Research Centre for Magnetic Resonance DRCMR, Section 714 Copenhagen University Hospital Amager and Hvidovre Kettegaard Allé 30, 2650 Hvidovre, Denmark
-- Torkil Svensgaard Sysadmin MR-Forskningssektionen, afs. 714 DRCMR, Danish Research Centre for Magnetic Resonance Hvidovre Hospital Kettegård Allé 30 DK-2650 Hvidovre Denmark Tel: +45 386 22828 E-mail: torkil@drcmr.dk
-- Alexander E. Patrakov