Distributed Compute Node deployment

Introduction

Additional groups of compute nodes can be deployed and integrated with an existing deployment of a control plane stack. These compute nodes are deployed in separate stacks from the main control plane (overcloud) stack, and they consume some of the stack outputs from the overcloud stack to reuse as configuration data.

Deploying these additional nodes in separate stacks provides for separation of management between the control plane stack and the stacks for additional compute nodes. The stacks can be managed, scaled, and updated separately.

Using separate stacks also creates smaller failure domains as there are less baremetal nodes in each invidiual stack. A failure in one baremetal node only requires that management operations to address that failure need only affect the single stack that contains the failed node.

A routed spine and leaf networking layout can be used to deploy these additional groups of compute nodes in a distributed nature. Not all nodes need to be co-located at the same physical location or datacenter. See Deploying Overcloud with L3 routed networking for more details.

Such an architecture is referred to as “Distributed Compute Node” or “DCN” for short.

Supported failure modes and High Availability recommendations

Handling negative scenarios for DCN starts from the deployment planning, like choosing some particular SDN solution over provider networks to meet the expected SLA.

Loss of control plane connectivity

A failure of the central control plane affects all DCN edge sites. There is no autonomous control planes at the edge. No OpenStack control plane API or CLI operations can be executed locally in that case. For example, you cannot create a snapshot of a Nova VM, or issue an auth token, nor can you delete an image or a VM.

Note

A single Controller service failure normally induces no downtime for edge sites and should be handled as for usual HA deployments.

Loss of an edge site

Running Nova VM instances will keep running. If stopped running, you need the control plane back to recover the stopped or crashed workloads.

Note

A single Compute service failure normally affects only its edge site without additional downtime induced for neighbor edge sites or the central control plane.

OpenStack infrastructure services, like Nova Compute, will automatically reconnect to MariaDB database cluster and RabbitMQ broker when the control plane’s uplink is back. No timed out operations can be resumed though and need to be retried manually.

It is recommended to maintain each DCN edge site as a separate Availability Zone (AZ) for Nova/Neutron and Cinder services.

Improving resiliency for N/S and E/W traffic

Reliability of the central control plane may be enhanced with L3 HA network, which only provides North-South routing. The East-West routing effectiveness of edge networks may be improved by using DVR or highly available Open Virtual Network (OVN). There is also BGPVPN and its backend specific choices.

Network recommendations

Traditional provider networks with backbone routing at the edge may fulfill or complement a custom distributed routing solution, like L3 Spine-Leaf topology.

Note

Neutron SDN backends that involve tunnelling may be sub-optimal for Edge DCN cases because of the known issues 1808594 and 1808062.

That said, when there is a network failure that disconnects the edge off the central site, there is no SLA for recovery time but only what the provider networks or a particular SDN choice can guarantee. For switched/routed/MPLS provider networks, that may span from 10’s of ms to a few seconds. With the outage thresholds are typically considered to be a 15 seconds. These trace back on various standards that are relevant here.

Config-drive/cloud-init details

The simplest solution we recommend for DCN would involve only provider networks at the edge. For that case, it is also recommended to use config-drive or another configuration mechanism other than cloud-init. Otherwise, the latter requires a 169.254.169.254/32 route for the provider routers to forward data to the metadata service.

IPv6 details

IPv6 for tenants’ workloads and infrastructure tunnels interconnecting the central site and the edge is a viable option as well. IPv6 cannot be used for provisioning networks though. Key benefits IPv6 may provide for DCN are:

  • SLAAC, which is a EUI-64 form of autoconfig that makes IPv6 addresses calculated based on MAC addresses and requires no DHCP services placed on the provider networks.
  • Improved mobility for endpoints, like NFV APIs, to roam around different links and edge sites without losing its connections and IP addresses.
  • End-to-end IPv6 has been shown to have better performance by large content networks. This is largely due to the presence of NAT in most end-to-end IPv4 connections that slows them down.

Storage recommendations

DCN with only ephemeral storage is available for Nova Compute services. That is up to the edge cloud applications to be designed to provide enhanced data availability, locality awareness and/or replication mechanisms.

Deploying from a centralized undercloud

The main overcloud control plane stack should be deployed as needed for the desired cloud architecture layout. This stack contains nodes running the control plane and infrastructure services needed for the cloud. For the purposes of this documentation, this stack is referred to as the overcloud stack.

The overcloud stack may or may not contain compute nodes. It may be a user requirement that compute services are available within the overcloud stack, however it is not strictly required.

Undercloud configuration

TODO

Saving configuration from the overcloud

Once the overcloud has been deployed, data needs to be retrieved from the overcloud Heat stack and plan to pass as input values into the separate DCN deployment.

Extract the needed data from the stack outputs:

# EndpointMap: Cloud service to URL mapping
openstack stack output show standalone EndpointMap --format json \
  | jq '{"parameter_defaults": {"EndpointMapOverride": .output_value}}' \
  > endpoint-map.json

# AllNodesConfig: Node specific hieradata (hostnames, etc) set on all nodes
openstack stack output show standalone AllNodesConfig --format json \
  | jq '{"parameter_defaults": {"AllNodesExtraMapData": .output_value}}' \
  > all-nodes-extra-map-data.json

# GlobalConfig: Service specific hieradata set on all nodes
openstack stack output show $STACK GlobalConfig --format json \
  | jq '{"parameter_defaults": {"GlobalConfigExtraMapData": .output_value}}' \
  > $DIR/global-config-extra-map-data.json

# HostsEntry: Entries for /etc/hosts set on all nodes
openstack stack output show standalone HostsEntry -f json \
  | jq -r '{"parameter_defaults":{"ExtraHostFileEntries": .output_value}}' \
  > extra-host-file-entries.json

The same passwords and secrets should be reused when deploying the additional compute stacks. These values can be saved from the existing control plane stack deployment with the following command:

.. code-block:: bash
openstack object save overcloud plan-environment.yaml python -c “import yaml; data=yaml.safe_load(open(‘plan-environment.yaml’).read()); print yaml.dump(dict(parameter_defaults=data[‘passwords’]))” > passwords.yaml

Use the passwords.yaml enviornment file generated by the previous command, or reuse the environment file used to set the values for the control plane stack.

Note

The passwords.yaml generated in the previous command contains sensitive security data such as passwords and TLS certificates that are used in the overcloud deployment.

Care should be taken to keep the file as secured as possible.

Create an environment file for setting necessary oslo messaging configuration overrides:

parameter_defaults:
  ComputeExtraConfig:
    oslo_messaging_notify_use_ssl: false
    oslo_messaging_rpc_use_ssl: false

Reusing networks from the overcloud

TODO

Spine and Leaf configuration

TODO

Standalone deployment

TODO