If we are talking about requests being blocked 60+ seconds, those tunings might not help (they help a lot for average latency during recovering/backfilling).
It would be interesting to see the logs for those blocked requests at OSD side (they have level 0), pattern to search might be "slow requests \d+ seconds old".
I had a problem that for a recovery candidate object, all updates to that object would stuck until it is recovered, that might take extremely long time if there are large number of PG and objects to recover. But I think that is resolved by Sam to allow write for degraded objects in Hammer.
----------------------------------------
> Date: Thu, 10 Sep 2015 14:56:12 -0600
> From: robert@xxxxxxxxxxxxx
> To: ceph-users@xxxxxxxxxxxxxx
> Subject: Hammer reduce recovery impact
>
> -----BEGIN PGP SIGNED MESSAGE-----
> Hash: SHA256
>
> We are trying to add some additional OSDs to our cluster, but the
> impact of the backfilling has been very disruptive to client I/O and
> we have been trying to figure out how to reduce the impact. We have
> seen some client I/O blocked for more than 60 seconds. There has been
> CPU and RAM head room on the OSD nodes, network has been fine, disks
> have been busy, but not terrible.
>
> 11 OSD servers: 10 4TB disks with two Intel S3500 SSDs for journals
> (10GB), dual 40Gb Ethernet, 64 GB RAM, single CPU E5-2640 Quanta
> S51G-1UL.
>
> Clients are QEMU VMs.
>
> [ulhglive-root@ceph5 current]# ceph --version
> ceph version 0.94.2 (5fb85614ca8f354284c713a2f9c610860720bbf3)
>
> Some nodes are 0.94.3
>
> [ulhglive-root@ceph5 current]# ceph status
> cluster 48de182b-5488-42bb-a6d2-62e8e47b435c
> health HEALTH_WARN
> 3 pgs backfill
> 1 pgs backfilling
> 4 pgs stuck unclean
> recovery 2382/33044847 objects degraded (0.007%)
> recovery 50872/33044847 objects misplaced (0.154%)
> noscrub,nodeep-scrub flag(s) set
> monmap e2: 3 mons at
> {mon1=10.217.72.27:6789/0,mon2=10.217.72.28:6789/0,mon3=10.217.72.29:6789/0}
> election epoch 180, quorum 0,1,2 mon1,mon2,mon3
> osdmap e54560: 125 osds: 124 up, 124 in; 4 remapped pgs
> flags noscrub,nodeep-scrub
> pgmap v10274197: 2304 pgs, 3 pools, 32903 GB data, 8059 kobjects
> 128 TB used, 322 TB / 450 TB avail
> 2382/33044847 objects degraded (0.007%)
> 50872/33044847 objects misplaced (0.154%)
> 2300 active+clean
> 3 active+remapped+wait_backfill
> 1 active+remapped+backfilling
> recovery io 70401 kB/s, 16 objects/s
> client io 93080 kB/s rd, 46812 kB/s wr, 4927 op/s
>
> Each pool is size 4 with min_size 2.
>
> One problem we have is that the requirements of the cluster changed
> after setting up our pools, so our PGs are really out of wack. Our
> most active pool has only 256 PGs and each PG is about 120 GB is size.
> We are trying to clear out a pool that has way too many PGs so that we
> can split the PGs in that pool. I think these large PGs is part of our
> issues.
>
> Things I've tried:
>
> * Lowered nr_requests on the spindles from 1000 to 100. This reduced
> the max latency sometimes up to 3000 ms down to a max of 500-700 ms.
> it has also reduced the huge swings in latency, but has also reduced
> throughput somewhat.
> * Changed the scheduler from deadline to CFQ. I'm not sure if the the
> OSD process gives the recovery threads a different disk priority or if
> changing the scheduler without restarting the OSD allows the OSD to
> use disk priorities.
> * Reduced the number of osd_max_backfills from 2 to 1.
> * Tried setting noin to give the new OSDs time to get the PG map and
> peer before starting the backfill. This caused more problems than
> solved as we had blocked I/O (over 200 seconds) until we set the new
> OSDs to in.
>
> Even adding one OSD disk into the cluster is causing these slow I/O
> messages. We still have 5 more disks to add from this server and four
> more servers to add.
>
> In addition to trying to minimize these impacts, would it be better to
> split the PGs then add the rest of the servers, or add the servers
> then do the PG split. I'm thinking splitting first would be better,
> but I'd like to get other opinions.
>
> No spindle stays at high utilization for long and the await drops
> below 20 ms usually within 10 seconds so I/O should be serviced
> "pretty quick". My next guess is that the journals are getting full
> and blocking while waiting for flushes, but I'm not exactly sure how
> to identify that. We are using the defaults for the journal except for
> size (10G). We'd like to have journals large to handle bursts, but if
> they are getting filled with backfill traffic, it may be counter
> productive. Can/does backfill/recovery bypass the journal?
>
> Thanks,
>
> - ----------------
> Robert LeBlanc
> PGP Fingerprint 79A2 9CA4 6CC4 45DD A904 C70E E654 3BB2 FA62 B9F1
> -----BEGIN PGP SIGNATURE-----
> Version: Mailvelope v1.0.2
> Comment: https://www.mailvelope.com
>
> wsFcBAEBCAAQBQJV8e5qCRDmVDuy+mK58QAAaIwQAMN5DJlhrZkqwqsVXaKB
> nnegQjG6Y02ObLRrg96ghHr+AGgY/HRm3iShng6E1N9CL+XjcHSLeb1JqH9n
> 2SgGQGoRAU1dY6DIlOs5K8Fwd2bBECh863VymYbO+OLgtXbpp2mWfZZVAkTf
> V9ryaEh7tZOY1Mhx7mSIyr9Ur7IxTUOjzExAFPGfTLP1cbjE/FXoQMHh10fe
> zSzk/qK0AvajFD0PR04uRyEsGYeCLl68kGQi1R7IQlxZWc7hMhWXKNIFlbKB
> lk5+8OGx/LawW7qxpFm8a1SNoiAwMtrPKepvHYGi8u3rfXJa6ZE38jGuoqRs
> 8jD+b+gS0yxKbahT6S/gAEbgzAH0JF4YSz+nHNrvS6eSebykE9/7HGe9W7WA
> HRAkrESi/f1MKtRkud2Nhycx2R0MZLK/HoumnCN8WUmgvOtKsyYpXj6FXghv
> VGpi3r6uyC5Xlb8JGREqB1hAUTHAv0+z4biDBvPYrENwFUaerWiIujIeLWV9
> aYuiQBjjDCLoqWZj0+gQwn9/zXo8gE7jo3XAemYqGB8NJY1e+RZW6+TgC2rD
> Floa1en1PzZsynm1Ho+RPWW509kog5fFkt41nJmmxRi3kNWwiJfKLJvysetl
> RYudFG1cEumfI68VyNcuL4dMzf9FsiADsBaHue8g9a5bjJH8LjK4fKZDCCJf
> Rzgu
> =vlrz
> -----END PGP SIGNATURE-----
> _______________________________________________
> 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
