I've been trying to understand how peer-to-peer networks work, and I can see how it works once you've discovered a peer (servant), but I completely fail to understand how you discover the initial host to connect to.

Can anyone enlighten me?

- Dale.

Right now, it is with the use of "helper services", better known as host caches. It is essentially a well known Gnutella peer that is always running - at least, they're supposed to according to the owners

Peers in need of more connections will connect to this particular host cache. The host cache will record the IP of the peer in its cache and send out a number (say 10) other IP addresses from previous peers. The client then uses those IP addresses to connect to another peer, and the cycle begins.

Such host caches do indeed introduce a "central" piece to a decentralized network, but they're not required part. They're just "helpers", to make your life a bit easier. If all these central host caches were removed, anyone could still connect to the network.

For instance, it could be as simple as your friends e-mailing you a specific e-mail address. It could be someone posting his IP address on a newsgroup or on a chat room.

You should also have added that actually now, there is no single central router node on the Gnutella network.
At least there are about 8 well-known routers from several countries and companies, and there are now users arounf the world that share their gnutella.net files to add further routers in their list of known routers.
What is a router ? just a standard Gnutella agent, that allows incoming connections, and makes some outgoing connections but that performs it only temporarily, just to collect peer adresses. Actually, for performance reason, a router does not perform searches and downloads, so that it allows more incoming connections just to send them some pong answers for a single ping request, or to provide you the addresses of other best-performing agents such as ultrapeers.
Ultrapeers are the best addition to the central router solution, as they collaborate efficiently to avoid centralizing much incoming traffic on very few routers.
LimeWire's Ultrapeer feature can involve thousands of well-connected Ultrapeers that can be automatically and dynamically discovered, while preserving the bandwidth and accessibility of the LimeWire's router. So the router does not have to actively browse the network to discover new hosts: any host entering to the network just needs a single connection to the router, and does not need any further connection to the router, as this will best work by using the highly-distributed Ultrapeer-capable agents.

But it's true: you need to give to the user-agent one or some initial address to connect to. This may be the address of an efficient router, or the address of an Ultrapeer, or the adress of any other agent. That's why the network cannot be disconnected.

Even the LimeWire's router does not know the adresses of each host on the network. It does not need it as there are always many ultrapeers in the network that manages that work for it.

The only way to shutdown the gnutella network would be to completely shutdown the whole Internet, which is nearly impossible. Even if this ever occured, due to a required global service migration, there would need to be some way to reconnect further, so that users can communicate on the Internet. As soon as this occurs, users will be able to communicate via IRC, online forums, Emails, and so on, so Ultrapeers will immediately restart their work in the new network.

Don't expect a major shutdown with a full technology change that forbids all legacy applications to work. At least the Emails services will be restored as it is the most wanted service on the Internet which is required by most (if not all) other applications, including those for the e-Commerce and advertizing area.