VNF scaling

VNF resources in terms of CPU core and memory are hardcoded in VNFD template through image flavor settings. This result in either provisioning VNF for typical usage or for maximum usage. The former leads to service disruption when load exceeds provisioned capacity. And the later leads to underutilized resources and waste during normal system load. So tacker provides a way to seamlessly scale the number of VNFs on demand either manually or automatically.

TOSCA schema for scaling policy

Tacker defines TOSCA schema for the scaling policy as given below:

tosca.policies.tacker.Scaling:
  derived_from: tosca.policies.Scaling
  description: Defines policy for scaling the given targets.
  properties:
    increment:
      type: integer
      required: true
      description: Number of nodes to add or remove during the scale out/in.
    min_instances:
      type: integer
      required: true
      description: Minimum number of instances to scale in.
    max_instances:
      type: integer
      required: true
      description: Maximum number of instances to scale out.
    default_instances:
      type: integer
      required: true
      description: Initial number of instances.
    cooldown:
      type: integer
      required: false
      default: 120
      description: Wait time (in seconds) between consecutive scaling
      operations. During the cooldown period, scaling action will be ignored
  targets:
    type: list
    entry_schema:
      type: string
    required: true
    description: List of Scaling nodes.

Sample TOSCA with scaling policy

Following TOSCA snippet shows the scaling policy used in VNFD, in which vdu1 and vdu2 are already defined VDUs.

policies:

  - sp1:

      type: tosca.policies.tacker.Scaling

      description: Simple VDU scaling

      targets: [vdu1, vdu2]

      properties:
        min_instances: 1

        max_instances: 3

        default_instances: 2

        increment: 1

Deploying scaling TOSCA template using Tacker

Once OpenStack/Devstack along with Tacker has been successfully installed, deploy a sample scaling template from location given below: https://github.com/openstack/tacker/tree/master/samples/tosca-templates/vnfd

Refer the ‘Getting Started’ link below on how to create a VNFD and deploy a VNF: https://docs.openstack.org/tacker/latest/install/getting_started.html

How to scale VNF using CLI

Tacker provides following CLI for scaling.

openstack vnf scale –vnf-id <vnf-id>

–vnf-name <vnf name> –scaling-policy-name <policy name> –scaling-type <type>

Here,

  • scaling-policy-name - Policy name defined in scaling VNFD

  • scaling-type - in or out

  • vnf-id - scaling VNF id

  • vnf-name - scaling VNF name

For example, to scale-out policy ‘sp1’ defined above, this cli could be used as below:

openstack vnf scale –vnf-name sample-vnf

–scaling-policy-name sp1 –scaling-type out

How to scale VNF using REST API

Tacker provides following REST API for scaling.

POST on v1.0/vnfs/<vnf-id>/actions

with body

{“scale”: { “type”: “<type>”, “policy” : “<scaling-policy-name>”}}

Here,

  • scaling-policy-name - Policy name defined in scaling VNFD

  • scaling-type - in or out

  • vnf-id - scaling VNF id

Response http status codes:

  • 202 - Accepted the request for doing the scaling operation

  • 404 - Bad request, if given scaling-policy-name and type are invalid

  • 500 - Internal server error, on scaling operation failed due to an error

  • 401 - Unauthorized

VNF state transitions during scaling operation

During the scaling operation, the VNF will be moving in below state transformations:

  • ACTIVE -> PENDING_SCALE_IN -> ACTIVE

  • ACTIVE -> PENDING_SCALE_IN -> ERROR

  • ACTIVE -> PENDING_SCALE_OUT -> ACTIVE

  • ACTIVE -> PENDING_SCALE_OUT -> ERROR

Limitations

Following features are not supported with scaling:

  • Auto-scaling feature is supported only with alarm monitors and it does not work with other monitors such as ping, http_ping.

  • When VNF is modelled with scaling requirement in VNFD, any config management requirement in VNFD is not supported.

  • Scaling feature does not support to selectively choose the VDU as part of scaling.