Packet switching against the circuit

1964Baran, at RAND

Two people arrived at packet switching independently and for different reasons, which is worth separating because the popular story merges them.

The first line of work was about survivability. A communications network with central switches fails when the switches are destroyed; a distributed mesh in which messages are divided into blocks and routed independently degrades instead of failing. That work was published as a series of reports and was explicitly about withstanding attack.

1965 and 1966Davies, at the National Physical Laboratory

The second line of work was about efficiency, and it produced the word. Interactive use of a computer over a line is extremely bursty: a person types, thinks, reads, types again, and the line carries nothing for the great majority of the time it is reserved.

Dividing traffic into small labelled units lets several conversations share one line by filling each other's gaps. The motivation was the cost of leased lines, not survival.

1960s and 1970sThe argument it actually won on

Circuit switching was not defeated on reliability. It was defeated on utilisation.

A reserved path is ideal for traffic that flows steadily, which speech does and interactive computing does not. When the load is bursts separated by silence, reservation wastes almost everything, and the waste is proportional to how conversational the use is.

widely repeatedThe popular version

The story that the network everyone now uses was designed to survive a nuclear exchange conflates the two lines of work.

Survivability was the motivation of one body of research. The network built afterwards was justified to its funders as a way of sharing scarce and expensive computers between institutions, and several of the people responsible have said plainly that survivability was not the goal. Both facts are true and they are about different projects.

1960s onwardsThe unit and why it is small

The size of a packet is a compromise nobody escapes. Large units carry more payload per header and waste more when one is lost, since the whole unit is resent. Small units waste a larger fraction on headers and multiply the per-unit work at every hop.

The sizes that settled are the residue of the link technologies of the time, and one of them, the figure of around fifteen hundred bytes, comes from the amount a particular early local network could transmit before its collision detection stopped working.

That figure is now the effective ceiling on almost every path on the internet, which means a limit derived from the electrical behaviour of a shared cable governs traffic on fibre that has nothing to do with it.

present dayWhat reservation still buys

Circuit thinking did not disappear; it returned as a feature. Reserved capacity, guaranteed paths and traffic classes all exist to give particular flows the property that packet switching deliberately abandoned.

They are sold as advanced capabilities and are, in substance, the older idea reintroduced where it was always better: steady traffic that cannot tolerate variable delay. The two designs were never right and wrong. One suits bursts and the other suits streams, and the network we have is the burst answer with stream features bolted on.

1974 onwardsWhat arrived at the ends instead

The machinery the network declined to provide was written once, carefully, and put in a transport protocol that almost everything uses: sequence numbers to restore order, acknowledgements and timers to recover loss, and a control loop to slow down when the path is congested.

The last of those is the interesting one. Nothing in the network tells a sender to slow down; the sender infers congestion from loss and delay. An entire discipline of traffic control therefore runs on inference about a network that reports nothing, which is the direct consequence of the choice this entry is about.

Why the ends were the right place

The argument that settled it is that a guarantee provided by the network can still be broken above it, so an application that needs certainty has to check anyway. If the check exists at the ends regardless, providing the same guarantee in the middle is duplicated work that every user pays for whether or not they need it.

That reasoning is why the network stayed simple and general, and why arguments for putting intelligence back into it recur every decade and are answered the same way each time.

present dayWhat the choice pushed to the edges

Packets arrive out of order, late, or not at all, because nothing reserved a path for them. Everything anybody wants from a connection has to be rebuilt at the ends: ordering, retransmission, pacing, and the illusion of a stream.

That relocation is the durable consequence. A network that promises little is simple, cheap and general, and every application on it carries the machinery to cope with what the network declined to guarantee. Both halves of that bargain are still being renegotiated in every new transport protocol.

What we cannot verify

Both lines of work are documented in published reports and can be read directly. Priority disputes and who influenced whom are argued about by participants and historians and we take no position. Accounts of what motivated the funding of the later network differ between participants, and the claim that it was purely about resource sharing is itself a simplification of a longer argument.