Fail over serves on different VLANS

[Simon Hobson]

Klaus Vink Slott <gqv604 at hum.ku.dk> wrote:

> We have been running 2 dhcp servers in fail over configuration for quite a while. Now we consider moving one server to a host in another building. This will also place the host on another VLAN.
> ...
> How does the server "see if the other server makes an offer".

It doesn't !

> Do they need to bee on the same VLAN or is it communicated directly between the servers?

It's direct.

OK, a bit of background.
If a client does not have a valid lease, or it knows the lease is not valid for the current network, it broadcasts a DHCP_discover packet. This should reach all DHCP servers servicing the network - either direct or via relay agents (helpers). All DHCP servers configured to service this client will respond with an offer, and the client will pick an offer - technically the client should pick the "best"* offer, in practice I think most clients just take the first that has everything they need.

For servers that made offers not taken up, they will simply cancel the offer once they time out and don't get a request from the client.

When a client already has a lease, it will renew (extend) it by unicasting a request to the server that gave it the lease. No other server will see this, and it doesn't need the help of any relay agents.

* For whatever definition of best it is using.


Note that I've specifically catered for an arbitrary number of servers - that's because you can have multiple servers servicing a network, with some restrictions to avoid problems. For example, you could run multiple servers with non-overlapping pools and that will work just fine.

With failover, the requirements are :
The two servers can communicate via unicast packets between them - to run the failover protocol.
The two servers can both service the network for DHCP - that means being able to receive (and reply to) broadcast packets either directly or via a helper, and to communicate via unicast packets with clients.

There's no requirement to be on the same network, or VLAN, or even the same site - as long as the connectivity requirements are met, location doesn't matter (though if on different continents, latency may have an effect !).

As an aside, the reason the servers will not automatically go into partner down state when communication with the partner is lost, is because the traffic may take different routes. Eg, servers at sites A and B may both be able to communicate with a client at site C - but not communicate with each other due to a network problem. An example might be if there are three leased lines, one is down, but there is no redundant routing via the third site. There are also more mundane causes - such as the servers using different interfaces for different traffic and a cable is unplugged, or a switch has failed, or ...