yahoo-eng-team team mailing list archive

[Matt Riedemann <mriedem.os@xxxxxxxxx>]

Public bug reported:

Chris Friesen has reported that in Newton with a live migration that has
~16 ports per instance, the "network-changed" events generated from
neutron when the vifs are unplugged from the source host can effectively
block the network info cache refresh that's called at the end of the
live migration operation. Details are in the IRC logs:

http://eavesdrop.openstack.org/irclogs/%23openstack-nova/%23openstack-
nova.2017-05-17.log.html#t2017-05-17T22:50:31

But this stands out:

cfriesen        mriedem: so it looks like _build_network_info_model()
costs about 200ms plus about 125ms per port since we query each port
separatly from neutron.  and the refresh_cache lock is held the whole
time

In Nova the 'network-changed' event is handled generically because there
is no port id in the event, so nova just refreshes the entire nw info
cache on the instance - which can be expensive and redundant since it's
doing a lot of queries to Neutron to build up information about ports,
fixed IPs, floating IPs, subnets and networks, and Neutron doesn't have
bulk query APIs or allow OR filters in the API for bulk queries on
things like floating IPs.

https://github.com/openstack/nova/blob/8d492c76d53f3fcfacdd945a277446bdfe6797b0/nova/compute/manager.py#L6854

Looking in neutron's code that sends the network-changed event, there is
a port in scope, it's just not sent like for network-vif-deleted events.

We should be able to scope the network-changed event to a specific port
on the neutron side and check for that on the nova side so we don't have
to refresh the entire network info cache, but just the vif that was
updated.

** Affects: nova
     Importance: Undecided
         Status: New

-- 
You received this bug notification because you are a member of Yahoo!
Engineering Team, which is subscribed to OpenStack Compute (nova).
https://bugs.launchpad.net/bugs/1691602

Title:
  live migration generates several network-changed events which lock up
  refreshing the nw info cache

Status in OpenStack Compute (nova):
  New

Bug description:
  Chris Friesen has reported that in Newton with a live migration that
  has ~16 ports per instance, the "network-changed" events generated
  from neutron when the vifs are unplugged from the source host can
  effectively block the network info cache refresh that's called at the
  end of the live migration operation. Details are in the IRC logs:

  http://eavesdrop.openstack.org/irclogs/%23openstack-nova/%23openstack-
  nova.2017-05-17.log.html#t2017-05-17T22:50:31

  But this stands out:

  cfriesen        mriedem: so it looks like _build_network_info_model()
  costs about 200ms plus about 125ms per port since we query each port
  separatly from neutron.  and the refresh_cache lock is held the whole
  time

  In Nova the 'network-changed' event is handled generically because
  there is no port id in the event, so nova just refreshes the entire nw
  info cache on the instance - which can be expensive and redundant
  since it's doing a lot of queries to Neutron to build up information
  about ports, fixed IPs, floating IPs, subnets and networks, and
  Neutron doesn't have bulk query APIs or allow OR filters in the API
  for bulk queries on things like floating IPs.

  https://github.com/openstack/nova/blob/8d492c76d53f3fcfacdd945a277446bdfe6797b0/nova/compute/manager.py#L6854

  Looking in neutron's code that sends the network-changed event, there
  is a port in scope, it's just not sent like for network-vif-deleted
  events.

  We should be able to scope the network-changed event to a specific
  port on the neutron side and check for that on the nova side so we
  don't have to refresh the entire network info cache, but just the vif
  that was updated.

To manage notifications about this bug go to:
https://bugs.launchpad.net/nova/+bug/1691602/+subscriptions