hostname for services vs. IP address

[Chris Hecker]

I want to have a UDP broadcast for the local subnet that looks for 
services on machines that won't necessarily have a hostname in any 
meaningful way (in other words, they'll have a hostname, but it's not 
going to be in any kind of DNS, so I'll be connecting by raw IP), and 
potentially a service directory that needs to be able to point to 
machines by IP address.  What's the right way to do service principals 
in this case?  Is service/129.168.1.5 a valid service principal?  If 
not, do I need to reply to queries about these machines with the 
hostname they think they are, so the client can ask the KDC for a ticket 
with that hostname?

More concretely:

- host a sends out a udp broadcast
- host b replies, currently with its listen ip address and port
- host a connects to b by ip:port

or, alternatively:

- host a contacts the directory
- the directory sends a a list of ip addresses and ports
- host a picks one and connects to ip:port

Now, how to kerberize this?  The directory is trivially kerberizable 
because it's at directory.example.com, but, what to return on queries, 
whether UDP broadcast or directory queries?  Do I have to return the 
ip:port:hostname, so the client can form service/hostname at REALM and ask 
the KDC for a ticket?

Thanks,
Chris