Port knocking is one of those server security topics that seems to come up every now and then, and when it does it always sparks a bit of debate over matters of practicality.
The idea behind port knocking is simple: The administrator of a server sets up a server with an Internet-accessible service. The administrator then closes all the ports on the server, including the ports that the service uses, and starts a daemon that monitors all incoming packets to the server. The daemon then opens the port corresponding to the service when and only when the daemon receives a series of packets on a specific set of ports in the correct order. The packets can be any type -- TCP SYNs, TCP ACKs, UDPs, whatever -- but they must be sent to the server on the correct ports and in the correct order to cause the daemon to open a service's port. Thus a port remains closed until someone "knocks" on the correct ports to cause the daemon to temporarily open it.
(Example: If the required sequence is TCP ACK packets on ports 333, 4444, and 55555, then the sequence of ACKs "27, 333, 4444, 55555, 42" would open the port, whereas the sequence "333, 27, 4444, 55555, 42" would not. I would tend to use TCP ACK packages, because they look the most boring to someone sniffing a network -- more below.)
Thus port knocking adds the security equivalent of another "password" to a service, because a client has to successfully knock on a server's ports in the correct order in order to open a port for business. Given 2^16 possible ports, about 4 possible protocols, and (usually) about 4 necessary correct port guesses, standard port knocking comes out to the equivalent of a ((2^16 * 4) ^ 4) = 2^72 bit key. This isn't too shabby a number, in terms of key size, especially if port knocking is just an extra layer of security for a service.
Thus begins the debate on the security practicality of using port knocking.
The main advantage of port knocking is that it conceals the very existence of a service until the port knock sequence is complete. An attacker cannot attack a service he cannot find, and until he finds out how to properly knock, every port on the host will appear closed to him. Thus port knocking is very helpful in situations where a server administrator wishes to conceal the very existence of a service. Service concealment was the primary motivation behind the development of port knocking.
Another advantage of port knocking is that it allows a port, when it is opened, to only be open for connection from a specific IP address. The port knocking daemon monitors all incoming packets and when it detects the correct "knock" it can not only open up a port but can open a port that accepts packets from the knocking IP only. Thus a WAN-side service can be told to accept connections only from certain IPs, but those IPs can be decided in real time.
Unfortunately, however, port knocking is a very fragile security policy. Since the knocking packets must always be sent in the clear it has the equivalent security of a password sent in cleartext. Anyone sniffing your network on either the client or server (or anyone who can trick the client into sending the knock sequence to a spoofed server, ie via DNS poisoning) end can tell what "knock" is used and replay the packets, effectively negating the security of port knocking. The good news is that unless an attacker is actively looking for a port knock sequence, the knock will look like normal boring network traffic. But if a sniffer is on the lookout for a knock sequence -- especially if they know which server it is destined to -- it's impossible to slip the knock past him.
Also, because port knocking requires the equivalent of a symmetric key problem, port knocking does not scale well to services that must handle many individual connections. The pork knocking "key" must be distributed securely to everyone who needs access to the server's service. Any cryptographer and/or security auditor will tell you that the symmetric key distribution infrastructure for this sort of thing is both annoying and brittle -- the larger the infrastructure to be maintained, the larger the hassle and the larger the potential for single point failure.
Obviously, scalability does not matter to the individual wishing to SSH into his home computer, but it is of major concern to anyone operating a server with more than four or five users, especially if users must be added, revoked, and have their "knocking keys" rotated. This is a classic (annoying) symmetric key distribution problem.
Port knocking is also impractical for popular/busy servers because the popularity of a server contradicts the very goal of port knocking: to conceal the very existence of a service on a server. If an attacker is aware of the nature of his target and knows (or at least has an idea of) what services the server has, he will not be satisfied if his port scan turns up empty. If he expects certain services to exist on the server, and if he is in any way persistent, he will investigate the server until he finds the services he knows exist, thus he will investigate the potential use of port knocking sooner or later. This does not result in instantaneous defeat of course, but the port knock is only a secure as a password sent in cleartext, which is a bad security measure to have to fall back on. The exact situation will dictate how easy the knock will be to sniff.
This quick assessment should make it obvious that the only practical use for port knocking is on small servers. Realistically, the service itself should be secure enough in both configuration and implementation to not require the additional security of port knocking -- it is an Internet service, after all -- but port knocking does add yet another level of plausible security, and the paranoid never underestimate defense in depth.
All things considered, port knocking is not too useful, despite being a fun idea. The only place I've actually observed it in use was by people at DefCon looking for any way to add any level of security to their home computers. But, obviously, those are the "I will because I can" types. Plus, when you're at DefCon you use anything to secure your home server that you can get. But outside of the experimental world, port knocking is only an interesting notion, it never sees wide-scale usage for a reason.
Note that Port knocking is not the same thing as Single Packet Authentication. Port knocking was the initial attempt to gain security by service invisibility. SPA is the more secure successor to port knocking, developed to address key problems in port knocking.