Improve logic that determines liveliness of dhcp agent

[Eugene Nikanorov]

It has been revealed that under load (when a bunch of networks is processed) DHCP agent fails to send heartbeat updates and neutron-server starts to consider it dead.

In network failover feature this has been worked around by checking if DHCP agent has just started up and giving it amount of time proportional to number of networks scheduled to it.

However in scheduling logic there's no such additional liveliness conditions, so starting agent may be considered dead and network will be scheduled on additional agent beyond configured limit.

So in order to avoid such negative side effect, agent liveliness check logic needs to be unified.

[yong sheng gong (gongysh)]

what do u mean by 'network failover feature'? where is the feature?

[OpenStack Infra (hudson-openstack)]

Fix proposed to branch: master
Review: https://review.openstack.org/152891

[OpenStack Infra (hudson-openstack)]

Reviewed: https://review.openstack.org/152891
Committed: https://git.openstack.org/cgit/openstack/neutron/commit/?id=28fe7e85cc26d1a0254990ebee1bececae06f374
Submitter: Jenkins
Branch: master

commit 28fe7e85cc26d1a0254990ebee1bececae06f374
Author: Eugene Nikanorov <email address hidden>
Date: Wed Feb 4 15:05:36 2015 +0300

    Unify logic that determines liveliness of DHCP agent

    For DHCP agents sometimes it's not enough to check agent's last heartbeat
    time because in its starting period the agent may fail to send state reports
    because it's busy processing networks.
    In rescheduling logic such DHCP agent is given additional time after start.
    Additional time is proportional to amount of networks the agent is hosting.
    Need to apply the same logic to DHCP agent scheduler to avoid a case
    when starting agent is considered dead and a network gets more hosting
    agents than configured.

    Change-Id: I0fe6244c7d2ed42e4744351be34f251318322c54
    Closes-Bug: #1417708