Install ceph

• Installation depends about which version you want as they‘re all locked into Fixed releases (argonaut, bobtail etc). So go here for install options for your distro and the version you want:
  • http://ceph.com/docs/master/install/

Video to ceph intro

https://www.youtube.com/watch?v=UXcZ2bnnGZg
http://www.youtube.com/watch?v=BBOBHMvKfyc&feature=g-high

Rebooting node stops everything / Set number of replicas across all nodes

This issue is more or less fixed in Cuttlefish+

Make sure that the min replica count is set to nodes-1.

ceph osd pool set <poolname> min_size 1

Then the remaing node[s] will start up with just 1 node if everything else is down.

Keep in mind this can potentially make stuff ugly as there are no replicas now.

More info here: http://comments.gmane.org/gmane.comp.file-systems.ceph.devel/10481

Add disks (OSD) or entire nodes

Use ceph-deploy tools from cuttlefish+ instead for all these sorts of things unless you know what you‘re doing (http://ceph.com/docs/master/rados/deployment/ )

Prepare the disk as usual (partition or entire disk) - format with filesystem of choosing. Add to fstab and mount. Add to /etc/ceph/ceph.conf and replicate the new conf to the other nodes.

Start the disk, I‘m assuming we‘ve added osd.12 (sdd) on ceph1 here.

## Prepare disk first, create partition and format it
<insert parted oneliner>
mkfs.xfs -f /dev/sdd1

## Create the disk
ceph osd create [uuid]

## Auth stuff to make sure that the OSD is accepted into the cluser:
mkdir /srv/ceph/[uuid_from_above]
ceph-osd -i 12 --mkfs --mkkey
ceph auth add osd.12 osd ‘allow *‘ mon ‘allow rwx‘ -i /etc/ceph/keyring.osd.12

## Start it
/etc/init.d/ceph start osd.12

## Add it to the cluster and allow replicated based on CRUSH map.
ceph osd crush set 12 osd.12 1.0 pool=default rack=unknownrack host=ceph1-osd

In the line above, if you exchange the pool/rack/host you can place your disk/node where you want.

If you add a new host entry, it will be the same as adding a new node (with the disk).

Check that is in the right place with:

ceph osd tree

More info here:

• http://ceph.com/docs/master/rados/operations/pools/
• http://ceph.com/docs/master/rados/operations/add-or-rm-osds/

Delete pools/OSD

Use ceph-deploy tools from cuttlefish+ instead for all these sorts of things unless you know what you‘re doing (http://ceph.com/docs/master/rados/deployment/ )

Make sure you have the right disk, run

ceph osd tree

to get an overview.

If OSD is member of a pool and/or active

Here we‘ll delete osd5

## Mark it out
ceph osd out 5

## Wait for data migration to complete (ceph -w), then stop it
service ceph -a stop osd.5

## Now it is marked out and down

Delete the OSD

## If deleting from active stack, be sure to follow the above to mark it out and down
ceph osd crush remove osd.5

## Remove auth for disk
ceph auth del osd.5

## Remove disk
ceph osd rm 5

## Remove from ceph.conf and copy new conf to all hosts

Add Monitor node/service

Use ceph-deploy tools from cuttlefish+ instead for all these sorts of things unless you know what you‘re doing (http://ceph.com/docs/master/rados/deployment/ )

Install ceph, add keys, ceph.conf, host files and prepare a storage for containing the maps.

Then add the monitor into the system (To keep quorum, keep either 1 or 3+ - not 2). Examples are adding monitor mon.21 with ip 192.168.0.68

cd /tmp; mkdir add_monitor; cd add_monitor
ceph auth get mon. -o key
> exported keyring for mon.

ceph mon getmap -o map
> got latest monmap

ceph-mon -i 21 --mkfs --monmap map --keyring key 
> ceph-mon: created monfs at /srv/ceph/mon21 for mon.21

ceph mon add 21 192.168.0.68
> port defaulted to 6789added mon.21 at 192.168.0.68:6789/0

/etc/init.d/ceph start mon.21

Add the info to the ceph.conf file:

[mon]
...
[mon.21]
    host = ceph2-mon
    mon addr = 192.168.0.68:6789
...

Replicating from OSD-based to replication across hosts in a ceph cluster

More info here: http://jcftang.github.com/2012/09/06/going-from-replicating-across-osds-to-replicating-across-hosts-in-a-ceph-cluster/

Replication - see current level pr. OSD

ceph osd dump

CRUSH maps

• Redistributing, [de]assembling and finetuning; more info here:
  • http://hpc.admin-magazine.com/Articles/RADOS-and-Ceph-Part-2

## save current crushmap in binary
ceph osd getcrushmap -o crush.running.map

## Convert to txt
crushtool -d crush.running.map -o crush.map

## Edit it and re-convert to binary
crushtool -c crush.map -o crush.new.map

## Inject into running system
ceph osd setcrushmap -i crush.new.map

## If you‘ve added a new ruleset and want to use that for a pool, do something like:
ceph osd pool default crush rule = 4

KVM - add disk

• Substitute <host name=‘192.168.0.67‘ port=‘6789‘/> with ip/port of your monitor[s]
• Make sure you have a distinct id for the slotNR: <address type=‘pci‘ domain=‘0x0000‘ bus=‘0x00‘ slot=‘0x04‘ function=‘0x0‘/>
• Generate new uuid: uuidgen
• Be sure to have added secret key to libvirt; detailed just below ( http://wiki.skytech.dk/index.php/Ceph_-_howto,_rbd,_lvm,_cluster#KVM_-_add_secret.2Fauth_for_use_with_ceph )
• Omit the <auth>..</auth> part if you‘re not using any auth

Pr host:

<disk type=‘network‘ device=‘disk‘>
      <driver name=‘qemu‘ type=‘raw‘/>
      <source protocol="rbd" name="test/disk2-qemu-5g:rbd_cache=1">
        <host name=‘192.168.0.67‘ port=‘6789‘/>
        <host name=‘192.168.0.68‘ port=‘6789‘/>
      </source>
      <auth username=‘admin‘ type=‘ceph‘>
        <secret type=‘ceph‘ uuid=‘7a91dc24-b072-43c4-98fb-4b2415322b0f‘/>
      </auth>
      <target dev=‘vdb‘ bus=‘virtio‘/>
      <address type=‘pci‘ domain=‘0x0000‘ bus=‘0x00‘ slot=‘0x04‘ function=‘0x0‘/>
    </disk>

Pr pool:

<pool type=‘rbd‘>
   <name>rbd</name>
   <uuid>f959641f-f518-4505-9e85-17d994e2a399</uuid>
   <source>
     <host name=‘192.168.0.67‘ port=‘6789‘/>
     <host name=‘192.168.0.68‘ port=‘6789‘/>
     <host name=‘192.168.0.69‘ port=‘6789‘/>
     <name>test</name>
     <auth username=‘admin‘ type=‘ceph‘>
       <secret type=‘ceph‘ uuid=‘7a91dc24-b072-43c4-98fb-4b2415322b0f‘/>
     </auth>
   </source>
</pool>

KVM - add secret/auth for use with ceph

If you‘re using any form of ceph auth, this needs to be added - else skip this part

Create a secret.xml file:

<secret ephemeral=‘no‘ private=‘no‘>
   <uuid>7a91dc24-b072-43c4-98fb-4b2415322b0f</uuid>
   <usage type=‘ceph‘>
     <name>admin</name>
   </usage>
</secret>

Use it:

virsh secret-define secret.xml
virsh secret-set-value 7a91dc24-b072-43c4-98fb-4b2415322b0f AQDAD8JQOLS9IxAAbox00eOmlM1h5ZLGPxHGHw==

The last key is the key from your /etc/ceph/keyring.admin

cat /etc/ceph/keyring.admin 
[client.admin]
        key = AQDAD8JQOLS9IxAAbox00eOmlM1h5ZLGPxHGHw==

Online resizing of KVM images (rbd)

Resize the desired block-image (here going from 30GB -> 40GB)

qemu-img resize -f rbd rbd:sata/disk3 40G
> Image resized.

Find the attached target device:

virsh domblklist rbd-test
> Target     Source
> ------------------------------------------------
> vdb        sata/disk2-qemu-5g:rbd_cache=1
> vdc        sata/disk3:rbd_cache=1
> hdc        -

Then use virsh to tell the guest that the disk has a new size:

virsh blockresize --domain rbd-test --path "vdc" --size 40G
> Block device ‘vdc‘ is resized

Check raw rbd info

rbd --pool sata info disk3
> rbd image ‘disk3‘:
>        size 40960 MB in 10240 objects
>        order 22 (4096 KB objects)
>        block_name_prefix: rb.0.13fb.23353e97
>        parent:  (pool -1)

Make sure you can see the change from dmesg (Guest should see the new size change).

dmesg
> [...]
> [75830.538557] vdb: detected capacity change from 118111600640 to 123480309760
> [...]

Then extend the partition - if it is a simple data volume, you can just fdisk, remove the old partition, create a new and access default values for start/end (Note, only applies to partitions which holds nothing else!)

ext3/ext4

Write the partition, fdisk -l to doublecheck the size, then remount the partition (the partition from above is mounted as a data dir under vdb1 in my case):

mount -o remount,rw /dev/vdb1

Check your fstab to make sure you get the correct options for the remount.

Afterwards, call the resize2fs:

resize2fs /dev/vdb1
> resize2fs 1.42.5 (29-Jul-2012)
> Filesystem at /dev/vdb1 is mounted on /home/mirroruser/mirror; on-line resizing required
> old_desc_blocks = 7, new_desc_blocks = 7
> The filesystem on /dev/vdb1 is now 28835584 blocks long.

doublecheck via df -h or the like.

XFS

xfs_growfs /share

data blocks changed from 118111600640 to 123480309760

All done

Add/move journal in running cluster

Before starting this, it will be a good idea to make sure the cluster does not automark OSD‘s out if they‘ve been down for default 300s. You can do that by issuing:

• ceph osd set nodown

I‘ll assume we want to update node#1 having OSD 0,1,2 to put a journal on SSD with 10GB. It currently reside on each OSD with a 512MB journal
Assuming we‘ll be mounting a SSD here: /srv/ceph/journal   -- This will then hold all journals as /srv/ceph/journal/osd$id/journal

A much better way is to create the above as partitions and not files and use those instead - I‘ll show inline:
# Relevant ceph.conf options
 -- existing setup --
[osd]
    osd data = http://www.mamicode.com/srv/ceph/osd$id> 
 
 If you set your cluster to not mark OSD‘s down, remember to remove it! 
 
 
 
  
 ceph osd unset nodown 
 
 
 
 
 Enjoy your new faster journal! 
 
 
 More info here (source of the above): 
 
 
 
  
 http://www.sebastien-han.fr/blog/2012/08/17/ceph-storage-node-maintenance/ 
 
 
 
 
 
 Prevent OSD from being marked down 
 
 
 This could be the case while adding new OSD/nodes and some osd keeps being marked down 
 
 
ceph osd set nodown
 
 
 Check it is set by issuing: 
 
 
ceph osd dump | grep flags
> flags no-down
 
 
 When done, remember to unset it or no OSD will ever get marked down! 
 
 
ceph osd unset nodown
 
 
 Doublecheck with ceph osd dump | grep flags 
 
 
 
 OSD optimizations 
 
 
 A few things have helped make my cluster more stable: 
 
 
 
 Increase osd timeout 
 
 
 I‘ve found the existing 30sec rule is sometimes too little for my 4disk lowend system. 
 
 
 Set the timeout accordingly (try and run debug first to determine if this is really the case). 
 
 
osd op thread timeout = 60
 
 
 
 Adjust transaction commit size 
 
 
 Default value is 300, which is a bit high for highly stressed system/lowcpu 
 
 
 It is adjustable inside the [osd] column: 
 
 
[osd]
...
    ## Added new transactional size as default value of 300 is too much
    osd target transaction size = 50
...
 
 
 
 Change number of disks which can be in recovery 
 
 
 Since my ‘cluster‘ is just 1 machine with 6 osd having multiple recoveries running in parallel will effectively kill it; so I‘ve adjusted it down from the default values: 
 
 
## Just 2 osd backilling kills a server
osd max backfills = 1
osd recovery max active = 1
 
 
 More info on all the various settings here: 
 
 
 
  
 http://ceph.com/docs/master/rados/configuration/osd-config-ref/#recovery 
 
  
 http://ceph.com/docs/master/rados/configuration/osd-config-ref/#backfilling 
 
 
 
 
 
 Adjust kernel values 
 
 
 Adjust nr_request in queue (staying in mem - default is 128) 
 
 
 
  
 http://www.monperrus.net/martin/scheduler+queue+size+and+resilience+to+heavy+IO 
 
 
 
 
echo 1024 > /sys/block/sdb/queue/nr_requests
 
 
 Change scheduler as pr. Inktank blogpost: 
 
 
 
  
 http://ceph.com/community/ceph-bobtail-performance-io-scheduler-comparison/ 
 
 
 
 
 
 Full cluster - how to recover 
 
 
 
  
 Add more disks 
 
  
 Try lower the weight ratio of the full disk 
 
  
 If everything fails, up the full ratio and immediately change the weight ratio of the disk 
   
 
    
 ceph pg set_full_ratio 0.98 
 
   
 
 
  
 If you‘re really desperate set change the replicate level down 
   
 
    
 ceph osd pool set <poolname> size existing_size-1 
 
    
 ceph osd pool set <poolname> min_size existing_min_size-1 
 
    
 Keep in mind this is very dangerous if you hit 1 as min_size; then if any disk fail you‘ll loose all data! 
 
   
 
 
 
 
 
 Alternatively, if you want to change the default values (0.85 as near_full, 0.95 as full), you can add this to the mon section of your ceph.conf: 
 
 
[mon]
...
    mon osd nearfull ratio = 0.92
    mon osd full ratio = 0.98
...
 
 
 
 Loosing 2 nodes in a 3 node system 
 
 
 Will cause the cluster to halt because of a new feature. Explained by Greg Farnum from Ceph mailinglist: 
 
 
This took me a bit to work out as well, but you‘ve run afoul of a new
post-argonaut feature intended to prevent people from writing with
insufficient durability. Pools now have a "min size" and PGs in that
pool won‘t go active if they don‘t have that many OSDs to write on.
The clue here is the "incomplete" state. You can change it with "ceph
osd pool foo set min_size 1", where "foo" is the name of the pool
whose min_size you wish to change (and this command sets the min size
to 1, obviously). The default for new pools is controlled by the "osd
pool default min size" config value (which you should put in the
global section). By default it‘ll be half of your default pool size.

So in your case your pools have a default size of 3, and the min size
is (3/2 = 1.5 rounded up), and the OSDs are refusing to go active
because of the dramatically reduced redundancy. You can set the min
size down though and they will go active.
 
 
 
 Import LVM image into rbd 
 
 
 Do give thought about which version of rbd to use! Format 2 gives additional [great] benefits over format 1 
 
 
 
  
 http://ceph.com/docs/master/man/8/rbd/#parameters 
 
  
 http://ceph.com/docs/master/man/8/rbd/#commands (check for "This requires format 2") 
 
 
 
 
 I‘ve succesfully snapshotted lots of running systems, but the catch is all disk io not still flushed will not make it. You can use something like fsfreeze to first freeze the guest filesystem (ext4, btrfs, xfs). It will basically flush the FS state to disk and blocking any future write access while maintaining Read accesses. 
 
 
 
  
 [fsfreeze] http://manpages.courier-mta.org/htmlman8/fsfreeze.8.html 
 
 
 
 
 
 
 
 
 Turned out to be really simple 
 
 
 I started by grabbing a snapshot of a running lvm 
 
 
lvcreate -L25G -s -n snap-webserver /dev/storage/webserver
 
 
 And then just feeding that snapshot directly into rbd: 
 
 
rbd import /dev/storage/snap-webserver sata/webserver

## Format 2:
rbd --format 2 import /dev/storage/snap-webserver sata/webserver

## Format 2 for cuttlefish+:
rbd --image-format 2 import /dev/storage/snap-webserver sata/webserver
 
 
 Here sata/webserver means pool sata. The webserver image will be automatically created. 
 
 
 
 Import LVM image via ssh to remote host 
 
 
## Using format 2 rbd images 
dd if=/dev/e0.0/snap-webhotel | pv | ssh root@remote-server ‘rbd --image-format 2 import - sata/webserver‘
 
 
 Will run and show total copied data as well as speed 
 
 
 
 Import ceph image to LVM 
 
 
 Is pretty much just the reverse: 
 
 
## Optional; create a snapshot first
rbd snap create sata/webserver@snap1

## Transfer image
rbd export sata/webserver@snap1 - | pv | ssh root@remote-server ‘dd of=/dev/lvm-storage/webserver‘
 
 
 
 Import from ceph cluster to another cluster 
 
 
 Prob. lots of ways to do it; I did it the usual way with import/export 
 
 
ssh root@remote-server ‘rbd export sata/webserver -‘ | pv | rbd --image-format 2 import - sata/webserver
 
 
 In both cases using ‘-‘ is using stdin. 
 
 
 
 Create pool with pg_num 
 
 
 It seems to be important to create pools which are big enough in has the number in ^ 2. 
 
 
 So 2048, 4096 etc. 
 
 
ceph osd pool create sata 4096
 
 
 
 Mapping RBD volumes with kernel rbd-fuse client 
 
 
 
 Mapping volumes 
 
 
 It is recommended to use a 3.6+ kernel for this! 
 
 
/sbin/modprobe rbd
/usr/bin/rbd map --pool sata webserver-data --id admin --keyring /etc/ceph/keyring.admin
 
 
 Then it will show up as device /etc/rbd[X] 
 
 
 Then you can format it, partition it or do what you want - and then mount it like a normal device. 
 
 
 
 Unmapping volumes 
 
 
 If you want to remove a mapping for a device, issue: 
 
 
## First unmount from system, then:
rbd unmap /dev/rbd[X]
 
 
 
 Flashcache 
 
 
 
  
 http://www.sebastien-han.fr/blog/2012/11/15/make-your-rbd-fly-with-flashcache/ 
 
  
 Options about how to use/mount the flashcache ( https://github.com/facebook/flashcache/blob/master/doc/flashcache-sa-guide.txt ) 
   
 
    
 Writethrough - safest, all writes are cached to ssd but also written to disk immediately. If your ssd has slower write performance than your disk (likely for early generation SSDs purchased in 2008-2010), this may limit your system write performance. All disk reads are cached (tunable). 
 
    
 Writearound - again, very safe, writes are not written to ssd but directly to disk. Disk blocks will only be cached after they are read. All disk reads are cached (tunable). 
 
    
 Writeback - fastest but less safe. Writes only go to the ssd initially, and based on various policies are written to disk later. All disk reads are cached (tunable). 
 
   
 
 
 
 
 
 
 Various stuff 
 
 
 
 Inflight IO 
 
 
 
  
 You can peek at in-flight IO (and other state) with 
   
 
    
 cat /sys/kernel/debug/ceph/*/osdc 
 
   
 
 
 
 
 
 
  
 Ask via socket admin call 
   
 
    
 ceph --admin-daemon /var/run/ceph/ceph-osd.NN.asok dump_ops_in_flight 
 
   
 
 
 
 
 
 
 NFS over rbd (with pacemaker in 2x setup) 
 
 
 
  
 http://www.sebastien-han.fr/blog/2012/07/06/nfs-over-rbd/ 
 
 
 
 
 
 Backup VM‘s using rbd (req. format 2) 
 
 
 
  
 http://www.wogri.at/Ceph-VM-Backup.339.0.html 
 
 
 
 
 
 Bring down a server/rack/whatever for maint and stop ceph from redistributing data 
 
 
 Source: http://ceph.com/docs/master/rados/operations/troubleshooting-osd/#stopping-w-out-rebalancing 
 
 
## Set cluster to now mark OSD down
ceph osd set noout

## Once the cluster is set to nodown, you can begin stopping the OSDs within the failure domain that requires maintenance work.
ceph osd stop osd.{num}

## Note Placement groups within the OSDs you stop will become degraded while you are addressing issues with within the failure domain.
## Once you have completed your maintenance, restart the OSDs.
ceph osd start osd.{num}

## Finally, you must unset the cluster from nodown.
ceph osd unset noout
 
 
 
 RBD 
 
 
 
 Create new rbd share in pool X 
 
 
 Using rbd 
 
 
 Creating a 50GB share in sata named rpmbuilder-ceph could be done like: 
 
 
rbd create sata/rpmbuilder-ceph [--format 2*] --size 51200
 
 
 also qemu-img can create rbd images; I fancy it works the same, but am not entirely sure. 
 
 
 \* This will default create format 1 rbd blockdevices; format 2 has to be specified and brings a lot of new utility handling images (flattening, cloning etc). Use --format 2 to create as format 2 images. 
 
 
 
 
 
 
 Using KVM 
 
 
qemu-img create -f rbd rbd:sata/rpmbuilder-ceph 50G
 
 
 
 QEMU / KVM 
 
 
 
 Stalled guests during recovery/calc. pg‘s 
 
 
 Note: Qemu 1.4.2+ has this patch incorporated, so if you‘re using that or anything newer, you‘re all set. 
 
 
 Is a known problem caused by qemu flushing synchronous ( http://git.qemu.org/?p=qemu.git;a=commit;h=dc7588c1eb3008bda53dde1d6b890cd299758155 ). 
 
 
 There is a patch ready here: http://git.qemu.org/?p=qemu.git;a=commitdiff;h=dc7588c1eb3008bda53dde1d6b890cd299758155 
 
 
 More on this from this thread: http://www.spinics.net/lists/ceph-users/msg01352.html 
 
 
 
 Compiling new qemu-kvm on ubuntu 13.04 patched with async patch 
 
 
 I suggest using a normal user to build 
 
 
 Install prereqs: 
 
 
## Get and install deps
sudo apt-get install build-essential checkinstall

## I‘m not sure if these are even needed
sudo apt-get install git-core mercurial

cd
mkdir build
cd build
sudo apt-get build-dep qemu-kvm

## Grab source and patch it
cd qemu-1.4.0+dfsg/debian/patches/
wget "http://git.qemu.org/?p=qemu.git;a=patch;h=dc7588c1eb3008bda53dde1d6b890cd299758155" -o 0020-async-rbd.patch
patch < 0020-async-rbd.patch

## Build it
dpkg-buildpackage -rfakeroot -b
 
 
 It will take some time to build. Afterwards you‘ll have a bunch of .deb packages in the builds dir. 
 
 
 Just install the qemu-kvm_1.4.0+dfsg-1expubuntu4_amd64.deb and you should be good to try out the new async patch. 
 
 
 
 Download patched kvm 
 
 
 If you don‘t want the hassle of compiling yourself and you trust complete strangers with .deb package you can download the one I made here (only amd-64 bit - and only 1.4.0 stock ubuntu 13.04). 
 
 
 http://utils.skarta.net/qemu-kvm/qemu-kvm_1.4.0+dfsg-1expubuntu4_amd64.deb 
 
 
 
 Manual repair of PG 
 
 
 In Cuttlefish and later the repair procedure is much more intelligent and should be able to safely find the correct pg - so it should be safe to use instead 
 
 
 This is copied from ceph-users mailingslist 
 
 
 First part of the mail: 
 
 
Some scrub errors showed up on our cluster last week. We had some issues with host stability a couple weeks ago; my guess is that errors were introduced at that point and a recent background scrub detected them. I was able to clear most of them via "ceph pg repair", but several remain. Based on some other posts, I‘m guessing that they won‘t repair because it is the primary copy that has the error. All of our pools are set to size 3 so there _ought_ to be a way to verify and restore the correct data, right?

Below is some log output about one of the problem PG‘s. Can anyone suggest a way to fix the inconsistencies?

2013-05-20 10:07:54.529582 osd.13 10.20.192.111:6818/20919 3451 : [ERR] 19.1b osd.13: soid 507ada1b/rb.0.6989.2ae8944a.00000000005b/5//19 digest 4289025870 != known digest 4190506501
2013-05-20 10:07:54.529585 osd.13 10.20.192.111:6818/20919 3452 : [ERR] 19.1b osd.22: soid 507ada1b/rb.0.6989.2ae8944a.00000000005b/5//19 digest 4289025870 != known digest 4190506501
2013-05-20 10:07:54.606034 osd.13 10.20.192.111:6818/20919 3453 : [ERR] 19.1b repair 0 missing, 1 inconsistent objects
2013-05-20 10:07:54.606066 osd.13 10.20.192.111:6818/20919 3454 : [ERR] 19.1b repair 2 errors, 2 fixed
2013-05-20 10:07:55.034221 osd.13 10.20.192.111:6818/20919 3455 : [ERR] 19.1b osd.13: soid 507ada1b/rb.0.6989.2ae8944a.00000000005b/5//19 digest 4289025870 != known digest 4190506501
2013-05-20 10:07:55.034224 osd.13 10.20.192.111:6818/20919 3456 : [ERR] 19.1b osd.22: soid 507ada1b/rb.0.6989.2ae8944a.00000000005b/5//19 digest 4289025870 != known digest 4190506501
2013-05-20 10:07:55.113230 osd.13 10.20.192.111:6818/20919 3457 : [ERR] 19.1b deep-scrub 0 missing, 1 inconsistent objects
2013-05-20 10:07:55.113235 osd.13 10.20.192.111:6818/20919 3458 : [ERR] 19.1b deep-scrub 2 errors
 
 
 Actual solution proposed from an inktank employeer: 
 
 
You need to find out where the third copy is.  Corrupt it.  Then let repair copy the data from a good copy.

$ ceph pg map 19.1b

You should see something like this:
osdmap e158 pg 19.1b (19.1b) -> up [13, 22, xx] acting [13, 22, xx]

The osd xx that is NOT 13 or 22 has the corrupted copy.    Connect to the node that has that osd.

Find in the mount for osd xx your object with name "rb.0.6989.2ae8944a.00000000005b"

$ find /var/lib/ceph/osd/ceph-xx -name ‘rb.0.6989.2ae8944a.00000000005b*‘ -ls
201326612    4 -rw-r--r--   1 root     root          255 May 22 14:11 /var/lib/ceph/osd/ceph-xx/current/19.1b_head/rb.0.6989.2ae8944a.00000000005b__head_XXXXXXXX__0

I would stop osd xx, first.  In this case we find the file is 255 bytes long.  In order to make sure this bad copy isn‘t used.  Let‘s make the file 1 byte longer.

$ truncate -s 256 /var/lib/ceph/osd/ceph-xx/current/19.1b_head/rb.0.6989.2ae8944a.00000000005b__head_XXXXXXXX__0

Restart osd xx.  Not sure how what command does that on your platform.

Verify that OSDs are all running.  Shows all osds are up and in.
$ ceph -s | grep osdmap
osdmap e6: 6 osds: 6 up, 6 in

$ ceph osd repair 19.1b
instructing pg 19.1b on osd.13 to repair
 
 
 
 Change read ahead buffer to improve read speeds 
 
 
 As read speeds go directly to the storage behind ssd or whatever you have it will be as slow as reading from all your distributed devices. 
 
 
 You should be able to push that performance a bit with changing 
 
 
echo 4096 > /sys/block/vda/queue/read_ahead_kb