Open Stack

On Fri, 2014-01-03 at 18:33 +0800, Li Ma wrote:
> > Right, the Neutron L3 agent is the only OpenStack service that is
> > not stateless, and therefore you cannot use traditional
> > load-balancing across a set of identical nodes.
> 
> Hi jay, how about dhcp-agent? I've went through the bps of neutron
> for icehouse release. There's no related topics on dhcp-agent.
> 
> I wonder how to implement the same functionality of dhcp-agent.

Hi Li, sorry for the delayed response!

Technically, the DHCP agent itself does not maintain any state. Rather,
the DHCP agent ensures that a DHCP server (dnsmasq) is running in the
Linux network namespace and providing IP addresses in the tenant's
subnet to the VM's vNIC MAC addresses. The dnsmasq process, running in
the network namespace, *does* save state, though. Unless something has
changed recently, I believe the default behavior of Neutron is to start
a dnsmasq process within the tenant's network namespace on only *one*
DHCP agent.

However, it is definitely possible to instruct Neutron to place a
dnsmasq process on multiple nodes with a DHCP agent servicing the node.
In fact, in the Chef OpenStack Network cookbook [1] there is a tool
written by Alan Meadows called neutron-ha-tool.py [2] that contains code
[3] that queries Neutron for a list of DHCP agents, and then ensures
that all DHCP agents provide a dnsmasq process for all networks. Check
out the script and you'll see what I mean. We run it in a cron job and
it ensures that if a node running a DHCP agent goes down, that there are
other nodes providing the tenant subnet with DHCP servicing.

All the best,
-jay

[1] https://github.com/stackforge/cookbook-openstack-network
[2]
https://github.com/stackforge/cookbook-openstack-network/blob/master/files/default/neutron-ha-tool.py
[3]
https://github.com/stackforge/cookbook-openstack-network/blob/master/files/default/neutron-ha-tool.py#L254