kolla-ansible/doc/ceph-guide.rst
Richa Yadav dce6fb1bab Fixed grammatical error in ceph-guide.rst
Made changes in section Managing Ceph to correct the grammatical mistake.

Signed-off-by: Richa Yadav <ricyadav@cisco.com>

Change-Id: I2ed13f766dd39a7ac816af86664986b8d2700d80
2016-02-25 14:54:27 -08:00

5.5 KiB

Ceph in Kolla

The out-of-the-box Ceph deployment requires 3 hosts with at least one block device on each host that can be dedicated for sole use by Ceph. However, with tweaks to the Ceph cluster you can deploy a "healthy" cluster with a single host and a single block device.

Requirements

  • A minimum of 3 hosts for a vanilla deploy
  • A minimum of 1 block device per host

Preparation and Deployment

To prepare a disk for use as a Ceph OSD you must add a special partition label to the disk. This partition label is how Kolla detects the disks to format and bootstrap. Any disk with a matching partition label will be reformatted so use caution.

To prepare an OSD as a storage drive, execute the following operations:

# <WARNING ALL DATA ON $DISK will be LOST!>
# where $DISK is /dev/sdb or something similar
parted $DISK -s -- mklabel gpt mkpart KOLLA_CEPH_OSD_BOOTSTRAP 1 -1

The following shows an example of using parted to configure /dev/sdb for usage with Kolla.

parted /dev/sdb -s -- mklabel gpt mkpart KOLLA_CEPH_OSD_BOOTSTRAP 1 -1
parted /dev/sdb print
Model: VMware, VMware Virtual S (scsi)
Disk /dev/sdb: 10.7GB
Sector size (logical/physical): 512B/512B
Partition Table: gpt
Number  Start   End     Size    File system  Name                      Flags
     1      1049kB  10.7GB  10.7GB               KOLLA_CEPH_OSD_BOOTSTRAP

Edit the [storage] group in the inventory which contains the hostname of the hosts that have the block devices you have prepped as shown above.

[storage]
controller
compute1

Enable Ceph in /etc/kolla/globals.yml:

enable_ceph: "yes"

RadosGW is optional, enable it in /etc/kolla/globals.yml:

enable_ceph_rgw: "yes"

RGW requires a healthy cluster in order to be successfully deployed. On initial start up, RGW will create several pools. The first pool should be in an operational state to proceed with the second one, and so on. So, in the case of an all-in-one deployment, it is necessary to change the default number of copies for the pools before deployment. Modify the file /etc/kolla/config/ceph.conf and add the contents:

[global]
osd pool default size = 1
osd pool default min size = 1

Finally deploy the Ceph-enabled OpenStack:

kolla-ansible deploy -i path/to/inventory

Using a Cache Tier

An optional cache tier can be deployed by formatting at least one cache device and enabling cache tiering in the globals.yml configuration file.

To prepare an OSD as a cache device, execute the following operations:

# <WARNING ALL DATA ON $DISK will be LOST!>
# where $DISK is /dev/sdb or something similar
parted $DISK -s -- mklabel gpt mkpart KOLLA_CEPH_OSD_CACHE_BOOTSTRAP 1 -1

Enable the Ceph cache tier in /etc/kolla/globals.yml:

enable_ceph: "yes"
ceph_enable_cache: "yes"
# Valid options are [ forward, none, writeback ]
ceph_cache_mode: "writeback"

After this run the playbooks as you normally would. For example:

kolla-ansible deploy -i path/to/inventory

Setting up an Erasure Coded Pool

Erasure code is the new big thing from Ceph. Kolla has the ability to setup your Ceph pools as erasure coded pools. Due to technical limitations with Ceph, using erasure coded pools as OpenStack uses them requires a cache tier. Additionally, you must make the choice to use an erasure coded pool or a replicated pool (the default) when you initially deploy. You cannot change this without completely removing the pool and recreating it.

To enable erasure coded pools add the following options to your /etc/kolla/globals.yml configuration file:

# A requirement for using the erasure-coded pools is you must setup a cache tier
# Valid options are [ erasure, replicated ]
ceph_pool_type: "erasure"
# Optionally, you can change the profile
#ceph_erasure_profile: "k=4 m=2 ruleset-failure-domain=host"

Managing Ceph

Check the Ceph status for more diagnostic information. The sample output below indicates a healthy cluster:

docker exec ceph_mon ceph -s
cluster 5fba2fbc-551d-11e5-a8ce-01ef4c5cf93c
 health HEALTH_OK
 monmap e1: 1 mons at {controller=10.0.0.128:6789/0}
        election epoch 2, quorum 0 controller
 osdmap e18: 2 osds: 2 up, 2 in
  pgmap v27: 64 pgs, 1 pools, 0 bytes data, 0 objects
        68676 kB used, 20390 MB / 20457 MB avail
              64 active+clean

If Ceph is run in an all-in-one deployment or with less than three storage nodes, further configuration is required. It is necessary to change the default number of copies for the pool. The following example demonstrates how to change the number of copies for the pool to 1:

docker exec ceph_mon ceph osd pool set rbd size 1

All the pools must be modified if Glance, Nova, and Cinder have been deployed. An example of modifying the pools to have 2 copies:

for p in images vms volumes backups; do docker exec ceph_mon ceph osd pool set ${p} size 2; done

If using a cache tier, these changes must be made as well:

for p in images vms volumes backups; do docker exec ceph_mon ceph osd pool set ${p}-cache size 2; done

The default pool Ceph creates is named 'rbd'. It is safe to remove this pool:

docker exec ceph_mon ceph osd pool delete rbd rbd --yes-i-really-really-mean-it