neutron: refreshing vif model for a server gets the same network multiple times

[Matt Riedemann]

As of this change in Rocky:

https://review.opendev.org/#/c/585339/

When refreshing the vif model for an instance, e.g. when we get a network-changed event with a specific port ID:

https://logs.opendev.org/26/674326/2/experimental/nova-osprofiler-redis/899a204/controller/logs/screen-n-cpu.txt.gz#_Aug_02_18_35_50_884613

Aug 02 18:35:50.884613 ubuntu-bionic-vexxhost-sjc1-0009723918 nova-compute[20428]: DEBUG nova.network.neutronv2.api [req-1a4c2dbf-df86-4044-a59f-f751a53c5ea6 req-b0e1e2f7-d126-4e55-909d-4803816ca80f service nova] [instance: 5bbe0419-fbeb-4667-8c56-785fdc1d0a62] Refreshing network info cache for port 252040d6-4469-46ec-88c3-85e599a43104 {{(pid=20428) _get_instance_nw_info /opt/stack/nova/nova/network/neutronv2/api.py:1756}}

We get the network for the port multiple times, first here:

https://github.com/openstack/nova/blob/600ecf3d9a5116d040cd18023ff270b91b06247d/nova/network/neutronv2/api.py#L2966

https://github.com/openstack/nova/blob/600ecf3d9a5116d040cd18023ff270b91b06247d/nova/network/neutronv2/api.py#L392

And then we pass that list of 1 network dict to _build_vif_model here:

https://github.com/openstack/nova/blob/600ecf3d9a5116d040cd18023ff270b91b06247d/nova/network/neutronv2/api.py#L2850

and pass it to _nw_info_build_network here:

https://github.com/openstack/nova/blob/600ecf3d9a5116d040cd18023ff270b91b06247d/nova/network/neutronv2/api.py#L2883

Which then calls _get_physnet_tunneled_info which gets the network again here:

https://github.com/openstack/nova/blob/600ecf3d9a5116d040cd18023ff270b91b06247d/nova/network/neutronv2/api.py#L1904

and/or here:

https://github.com/openstack/nova/blob/600ecf3d9a5116d040cd18023ff270b91b06247d/nova/network/neutronv2/api.py#L1928

Furthermore, when we're doing forced _heal_instance_info_cache (stein+) we'll refresh the vif model for all ports that are currently attached to the server:

https://github.com/openstack/nova/blob/600ecf3d9a5116d040cd18023ff270b91b06247d/nova/network/neutronv2/api.py#L3015

And rebuild the vif model per port here:

https://github.com/openstack/nova/blob/600ecf3d9a5116d040cd18023ff270b91b06247d/nova/network/neutronv2/api.py#L3027

If there is more than one port on the same network attached to the server, we'll be calling show_network for each port even though we're getting the same data when those ports are on the same network.

I noticed this while checking some osprofiler results and noticed the network-changed event on the port-targeted refresh took a relatively long time:

https://logs.opendev.org/26/674326/2/experimental/nova-osprofiler-redis/899a204/osprofiler-traces/trace-fc50ca23-a6c2-474a-ac07-e61e706eb27d.html.gz

[Matt Riedemann (mriedem)]

There are similar inefficiencies here:

https://github.com/openstack/nova/blob/600ecf3d9a5116d040cd18023ff270b91b06247d/nova/network/neutronv2/api.py#L2014

https://github.com/openstack/nova/blob/600ecf3d9a5116d040cd18023ff270b91b06247d/nova/network/neutronv2/api.py#L2031

The former in the case we have multiple ports on the same network, the latter if we have a server create request with multiple instances of the same network (the user is telling nova to create multiple ports on the same network).

[OpenStack Infra (hudson-openstack)]

Fix proposed to branch: master
Review: https://review.opendev.org/674422

[OpenStack Infra (hudson-openstack)]

Change abandoned by Matt Riedemann (<email address hidden>) on branch: master
Review: https://review.opendev.org/674422
Reason: I'm no longer working on this. There are some ideas in the comments to simplify but someone else would have to take that on.