Generalize the Network Data Model

Articles » Network Automation Data Model Optimization » Generalize the Network Data Model

After starting with simplistic data model full of duplicate data, removing duplicate data, and restructuring the data model into a network-focused one we’re pretty close to an ideal solution. The only thing that still annoys me is the left and right part of an edge (link). In an undirected graph, we shouldn’t differentiate between the two ends of a connection. Furthermore, our data model can’t cope with rare multi-access links we might still find in real-life networks (examples: Carrier Ethernet E-LAN, DMVPN tunnels, wireless…).

At this point, we’re slightly diverging from “network is a graph” paradigm. Modeling multi-access network as a graph requires an extra node representing the network… in case you ever wondered why OSPF needs type-2 LSAs ;)

We can solve both challenges by replacing the left/right attributes of a link with a dictionary of nodes. Our revised data model would look like this:

Data structure describing all links in our fabric
links:
#
# Core link connecting S1 and S2
#
- S1:
    GigabitEthernet0/1: 172.16.0.1/30
  S2:
    GigabitEthernet0/1: 172.16.0.2/30

#
# Edge (stub) links on S1 and S2
#
- S1:
    Vlan101: 192.168.1.1/24
- S2:
    Vlan101: 192.168.2.1/24

Not only have we made our data model cleaner, but it also allows us to express corner cases like:

  • multi-access interfaces: a dictionary of nodes has more than two elements;
  • stub interfaces: a dictionary of nodes has just one element.

The only parameters left in the per-node host_vars files would be the node name and the AS number.

Host variables for S1
---
hostname: S1
bgp_as: 65001

Generating Configuration from Revised Data Model

Revised data model uses the same data structure for point-to-point and multi-access transit links as well as stub interfaces, resulting in simpler Jinja2 templates. All we have to do is to traverse the whole data structure:

  • Iterate over all links
  • Iterate over all nodes within a link data structure and perform some action when encountering the local node name
  • Iterate over all interfaces within the local node name

Here’s the part of the Jinja2 template that generates interface configuration:

{% macro interface(name,addr) -%}
!
interface {{ name }}
 ip address {{ addr|ipaddr('address') }} {{ addr|ipaddr('netmask') }}
{%- endmacro %}
{#
  Interfaces
#}
{% for link in links %}
{%   for node,iflist in link|dictsort if node == inventory_hostname %}
{%     for ifname,ip in iflist|dictsort %}
{{ interface(ifname,ip) }}
{%     endfor %}
{%   endfor %}
{% endfor %}

The template generating BGP network statements is almost identical to the interface code. All we need is an additional check: we need the network statements only for links that have a single node:

router bgp {{ bgp_as }}
...
{% for link in links if link|length == 1 %}
{% for node,iflist in link|dictsort if node == inventory_hostname %}
{% for ifname,ip in iflist|dictsort %}
 network {{ ip|ipaddr('network') }} {{ ip|ipaddr('netmask') }}
{% endfor %}
{% endfor %}
{% endfor %}

The BGP neighbor template is a bit more convoluted: we have to check whether the link is connected to the local node, and create neighbor statements for all other nodes connected to the same subnet:

{#
  BGP neighbor macro
#}
{% macro neighbor(name,ip) -%}
{% set n_ip = ip|ipaddr('address') %}
 neighbor {{ n_ip }} remote-as {{ hostvars[name].bgp_as }}
 neighbor {{ n_ip }} description {{ name }}
{%- endmacro %}
{#
  BGP routing protocol configuration
#}
router bgp {{ bgp_as }}
{#
  BGP neighbors - find links that contain local nodename and create
  neighbors for all other nodenames
#}
{% for link in links if link|length > 1 and inventory_hostname in link.keys() %}
{% for node,iflist in link|dictsort if node != inventory_hostname %}
{% for ifname,ip in iflist|dictsort %}
{{ neighbor(node,ip) }}
{% endfor %}
{% endfor %}
{% endfor %}

Not too bad, right? Can we do better? How about getting rid of IP addresses? You might want to try to do that step yourself before moving on.

More Information

Sidebar