I'll try to give here the most wide range, up to date and accurate information related to multicasting over TCP/IP networks that I can. Any feedback is very welcome. If you find any mistakes in this document, have any comments about its contents or an update or addition, please send them to me at the address listed at the top of this howto.
Multicast is... a need. Well, at least in some scenarios. If you have information (a lot of information, usually) that should be transmitted to various (but usually not all) hosts over an internet, then Multicast is the answer. One common situation in which it is used is when distributing real time audio and video to the set of hosts which have joined a distributed conference.
Multicast is much like radio or TV in the sense that only those who have tuned their receivers (by selecting a particular frequency they are interested on) receive the information. That is: you hear the channel you are interested in, but not the others.
Unicast is anything that is not broadcast nor multicast. All right, the definition is not very bright... When you send a packet and there is only one sender process -yours- and one recipient process (the one you are sending the packet to), then this is unicast. TCP is, by its own nature, unicast oriented. UDP supports a lot more paradigms, but if you are sending UDP packets and there is only one precess supposed to receive them, this is unicast too.
For years unicast transmissions proved to be enough for the Internet. It was not until 1993 when the first implementation of multicast saw the light in the 4.4 BSD release. It seems nobody needed it until then. Which were those new problems that multicast addressed?
Needless to say that the Internet has changed a lot since the "early days". Particularly, the appearance of the Web strongly transformed the situation: people didn't just want connections to remote hosts, mail and FTP. First they wanted to see the pictures people placed in their home pages, but later they also wanted to see and hear that people.
With today's technology it is possible to afford the "cost" of making a unicast connection with everyone who wants to see your web page. However, if you are to send audio and video, which needs a huge amount of bandwidth compared with web applications, you have -you had, until multicast came into scene- two options: to establish a separate unicast connection with each of the recipients, or to use broadcast. The first solution is not affordable: if we said that a single connection sending audio/video consumes a huge bandwidth, imagine having to establish hundreds or, may be, thousands of those connections. Both the sending computer and your network would collapse.
Broadcast seems to be a solution, but it's not certainly the solution. If you want all the hosts in your LAN to attend the conference, you may use broadcast. Packets will be sent only once and every host will receive them as they are sent to the broadcast address. The problem is that perhaps only a few of the hosts and not all are interested in those packets. Furthermore: perhaps some hosts are really interested in your conference, but they are outside of your LAN, a few routers away. And you know that broadcast works fine inside a LAN, but problems arise when you want broadcast packets to be routed across different LANs.
The best solution seems to be one in which you send packets to a certain special address (a certain frequency in radio/TV transmissions). Then, all hosts which have decided to join the conference will be aware of packets with that destination address, read them when they traverse the network, and pass them to the IP layer to be demultiplexed. This is similar to broadcasting in that you send only one broadcast packet and all the hosts in the network recognize and read it; it differs, however, in that not all multicast packets are read and processed, but only those that were previously registered in the kernel as being "of interest".
Those special packets are routed at kernel level like any packet because they are IP packets. The only difference might reside in the routing algorithm which tells the kernel where to route or not to route them.