Controlling Node Placement and IP Assignment

By default, nodes are assigned randomly via the Nova scheduler, either from a generic pool of nodes, or from a subset of nodes identified via specific profiles which are mapped to Nova flavors (See Baremetal Environment and Node matching with resource classes and profiles for further information).

However in some circumstances, you may wish to control node placement more directly, which is possible by combining the same capabilities mechanism used for per-profile placement with per-node scheduler hints.

Assign per-node capabilities

The first step is to assign a unique per-node capability which may be matched by the Nova scheduler on deployment.

This can either be done via the nodes json file when registering the nodes, or alternatively via manual adjustment of the node capabilities, e.g:

baremetal node set <id> --property capabilities='node:controller-0'

This has assigned the capability node:controller-0 to the node, and this must be repeated (using a unique continuous index, starting from 0) for all nodes.

If this approach is used, all nodes for a given role (e.g Controller, Compute or each of the Storage roles) must be tagged in the same way, or the Nova scheduler will be unable to match the capabilities correctly.

Note

Profile matching is redundant when precise node placement is used. To avoid scheduling failures you should use the default “baremetal” flavor for deployment in this case, not the flavors designed for profile matching (“compute”, “control”, etc).

Create an environment file with Scheduler Hints

The next step is simply to create a heat environment file, which matches the per-node capabilities created for each node above:

parameter_defaults:
    ControllerSchedulerHints:
        'capabilities:node': 'controller-%index%'

This is then passed via -e scheduler_hints_env.yaml to the overcloud deploy command.

The same approach is possible for each role via these parameters:

  • ControllerSchedulerHints

  • ComputeSchedulerHints

  • BlockStorageSchedulerHints

  • ObjectStorageSchedulerHints

  • CephStorageSchedulerHints

For custom roles (defined via roles_data.yaml) the parameter will be named RoleNameSchedulerHints, where RoleName is the name specified in roles_data.yaml.

Note

Previously the parameter for Compute nodes was named NovaComputeSchedulerHints. If you are updating a deployment which used the old parameter, all values previously passed to NovaComputeSchedulerHints should be passed to ComputeSchedulerHints instead, and NovaComputeSchedulerHints: {} should be explicitly set in parameter_defaults, to ensure that values from the old parameter will not be used anymore.

Custom Hostnames

In combination with the custom placement configuration above, it is also possible to assign a specific baremetal node a custom hostname. This may be used to denote where a system is located (e.g. rack2-row12), to make the hostname match an inventory identifier, or any other situation where a custom hostname is desired.

To customize node hostnames, the HostnameMap parameter can be used. For example:

parameter_defaults:
  HostnameMap:
    overcloud-controller-0: overcloud-controller-prod-123-0
    overcloud-controller-1: overcloud-controller-prod-456-0
    overcloud-controller-2: overcloud-controller-prod-789-0
    overcloud-novacompute-0: overcloud-novacompute-prod-abc-0

The environment file containing this configuration would then be passed to the overcloud deploy command using -e as with all environment files.

Note that the HostnameMap is global to all roles, and is not a top-level Heat template parameter so it must be passed in the parameter_defaults section. The first value in the map (e.g. overcloud-controller-0) is the hostname that Heat would assign based on the HostnameFormat parameters. The second value (e.g. overcloud-controller-prod-123-0) is the desired custom hostname for that node.

Predictable IPs

For further control over the resulting environment, overcloud nodes can be assigned a specific IP on each network as well. This is done by editing environments/ips-from-pool-all.yaml in tripleo-heat-templates. Be sure to make a local copy of /usr/share/openstack-tripleo-heat-templates before making changes so the packaged files are not altered, as they will be overwritten if the package is updated.

The parameter_defaults section in ips-from-pool-all.yaml, is where the IP addresses are assigned. Each node type has an associated parameter - ControllerIPs for Controller nodes, ComputeIPs for Compute nodes, etc. Each parameter is a map of network names to a list of addresses. Each network type must have at least as many addresses as there will be nodes on that network. The addresses will be assigned in order, so the first node of each type will get the first address in each of the lists, the second node will get the second address in each of the lists, and so on.

For example, if three Ceph storage nodes were being deployed, the CephStorageIPs parameter might look like:

CephStorageIPs:
  storage:
  - 172.16.1.100
  - 172.16.1.101
  - 172.16.1.102
  storage_mgmt:
  - 172.16.3.100
  - 172.16.3.101
  - 172.16.3.102

The first Ceph node would have two addresses: 172.16.1.100 and 172.16.3.100. The second would have 172.16.1.101 and 172.16.3.101, and the third would have 172.16.1.102 and 172.16.3.102. The same pattern applies to the other node types.

Important

Even if an overcloud node is deleted, its entry in the IP lists should not be removed. The IP list is based on the underlying Heat indices, which do not change even if nodes are deleted. To indicate that a given entry in the list is no longer used, the IP value can be replaced with a value such as “DELETED” or “UNUSED”.

In short, entries should never be removed from the IP lists, only changed or added.

To apply this configuration during a deployment, pass the environment file to the deploy command. For example, if you copied tripleo-heat-templates to ~/my-templates, the extra parameter would look like:

-e ~/my-templates/environments/ips-from-pool-all.yaml

Predictable Virtual IPs

You can also assign predictable Virtual IPs (VIPs) for services. To accomplish this, edit the network environment file and add the VIP parameters in the parameter_defaults section, for example:

ControlFixedIPs: [{'ip_address':'192.168.201.101'}]
InternalApiVirtualFixedIPs: [{'ip_address':'172.16.0.9'}]
PublicVirtualFixedIPs: [{'ip_address':'10.1.1.9'}]
StorageVirtualFixedIPs: [{'ip_address':'172.16.1.9'}]
StorageMgmtVirtualFixedIPs: [{'ip_address':'172.16.3.9'}]
RedisVirtualFixedIPs: [{'ip_address':'172.16.0.8'}]

These IPs MUST come from outside their allocation range to prevent conflicts. Do not use these parameters if deploying with an external load balancer.