Circuits versus packets
Traditional telephone networks work by circuit switching. When you place a call, the network sets up a path between you and the other person and holds it for the length of the call, even during silences. That works well for voice but poorly for computers, which tend to send data in bursts separated by long idle periods. Holding a line open for a computer that is mostly silent wastes capacity.
Packet switching takes a different approach. A message is divided into small pieces, each labeled with its destination and carried separately through the network. The pieces share the lines with other traffic, and a receiving computer puts them back in order. Because many users divide up the capacity, the network is used more efficiently.
Who developed the idea
Credit for packet switching is shared. Paul Baran of the RAND Corporation described a distributed network design in a series of reports from 1960 to 1964, in which messages were broken into blocks and forwarded through many redundant routes. His concern was building communication systems that could survive damage. Donald Davies at Britain’s National Physical Laboratory arrived at a similar idea independently in the mid-1960s, and he coined the word packet. Leonard Kleinrock had earlier worked out mathematical results about queues in networks.
Bob Taylor and Larry Roberts at ARPA chose packet switching for the ARPANET plan in 1967 and 1968, after Roberts learned of Davies’s work and Baran’s reports. Davies’s group at NPL built a small packet network in Britain, and in France Louis Pouzin’s Cyclades network, developed in the early 1970s, explored a simpler design that left more of the work to the computers at the edges. That design, in which individual packets called datagrams travel independently, strongly influenced the later internet.
The problem of many networks
By the early 1970s, there was not one network but several, and they differed. ARPANET used leased telephone lines. A packet radio network in the San Francisco area sent data over the air, and an experimental network using communications satellites spanned the Atlantic. Each had its own packet sizes, error behavior and addressing. A user on one could not exchange data with a user on another.
Vinton Cerf, then at Stanford, and Robert Kahn, at ARPA, set out to solve this. They published a design in 1974 for a protocol that would work across network boundaries. The key idea was to put minimal demands on the networks themselves. Each network would only need to deliver packets on a best-effort basis; the computers at the ends would handle lost or reordered data. Special machines called gateways, now called routers, would connect the networks.
How TCP/IP divides the work
The first design combined everything in one protocol, but experience showed it was better to separate two jobs. The Internet Protocol, IP, handles addressing and forwarding: it gets each packet one hop at a time toward the right machine, with no promises. The Transmission Control Protocol, TCP, runs on the end computers. It numbers the data, checks that everything arrived, asks for anything missing and slows down when the network is crowded.
This split is one reason the system has lasted. IP is simple enough to run on almost any underlying technology. TCP, or other protocols placed above IP, can be changed without altering the routers in between. Jon Postel at the University of Southern California, who edited the numbered documents (RFCs) that recorded these standards, became a central figure in keeping them consistent.
Switching over
Testing went on through the late 1970s, and ARPANET moved from its older protocol to TCP/IP on January 1, 1983. After that date, any network that implemented the protocols could join. Addressing needed to scale too, and in the mid-1980s the Domain Name System allowed people to use names rather than numbers.
The original addresses had 32 bits, which gives a little over four billion combinations, since 2 to the 32nd power equals 4,294,967,296. That seemed generous in the early 1980s, but the supply of unused addresses ran short as the network grew. A newer version, IPv6, uses much longer 128-bit addresses and has been deployed gradually since the late 1990s, alongside the older version.
Rival approaches
TCP/IP was not the only candidate. Telephone administrations promoted X.25, which resembled a virtual circuit and put more responsibility inside the network. International standards bodies developed a reference model for open systems and a full protocol suite based on it. In practice, TCP/IP won out because working software was available, often at no cost, and because it ran on the systems universities already used, such as Berkeley’s version of Unix.
What it made possible
A network built from packets and open protocols could grow without central planning. Anyone could attach new equipment if it followed the standards. This is also why the same underlying design carries email, web pages, video calls and software updates, and why the technology has been adapted from the slow leased lines of the 1970s to optical fiber and mobile cellular networks.