THE INTERNET: PAST, PRESENT, AND FUTURE The Beginnings of Packet Switching: Some Underlying Concepts Paul Baran, Com21, Inc. ABSTRACT volume to the matter of history and alternative approaches considered [3], which I recommend This article was written for a seminar held on to the serious reader to understand where this the occasion of the Franklin Institute’s 2001 work stood in relation to other work at the time. Bower Award and Prize for the Achievement in This activity was undertaken in 1960 at the Science to the author “for his seminal invention RAND Corporation. RAND was established by of packet switching — the foundation of modern the U.S. Air Force to preserve the operations communications networks and, in particular the research capability created by the Air Force in Internet.” World War II, and to work on issues of national It describes the author’s work 40 years ago security. The freedom of the staff to choose pro- focusing on the rationale creating the key con- jects, try novel approaches, and disagree with the cepts of packet switching. The article considers bureaucracy along the way is difficult to imagine the development of each of a series of about 20 in the present environment. Today, proposals essential concepts. For example, it examines must be written, projects excessively monitored, such subjects as the degree of redundancy to and reports prepared whether or not there is achieve any desired level of survivability; the anything worthwhile to report.1 It was a different necessity to chop data streams into small blocks; era then, and the remarkable degree of freedom what information had to be appended to these I enjoyed that encouraged farout, and sometimes blocks to allow the each block to find its own wild, thinking would be hard to duplicate today. way through the network; why it was necessary for each network element to operate at an inde- COLD WAR BACKGROUND pendent data rate; why all signals had to be con- verted to digital, and so on. When I joined RAND in 1959, a glaring weak In brief it describes the “why” as well as the spot in our strategic forces command and control “how” of packet switching works. communications was a dependence on shortwave radio and the national telephone system, AT&T, INTRODUCTION both highly vulnerable to attack. H-bomb testing in the Pacific revealed that long distance short- This article is a nostalgic trip, revisiting work I wave (high-frequency) sky-wave transmission did 40 years ago. It describes why packet switch- would be disrupted for several hours by a high- ing was invented, how it works, and why it seems altitude nuclear blast. Computer simulations to work so well. A personal recollection such as showed that weapons targeted at U.S. retaliatory this tends to overemphasize the importance of forces would render long distance telephone one’s own work. Throughout the design process, communications service inoperative by collateral I borrowed freely using whatever technology best damage alone. While most of the telephone fit the objective. I mention this to avoid any facilities would survive, the paucity of switching 1 I am particularly indebt- inadvertent impression that all the ideas to be centers formed a dangerous Achilles’ heel. ed to Frank Colbohm, described are totally original. Furthermore, my To cool tensions at this stage of the cold war, John D. Williams, Frank work focuses on basic concepts, not implementa- a retaliatory force capability was needed that Eldridge, Albert Wohlstet- tion — which was done by others. could withstand a surprise attack, and survive ter, Paul Armer, Willis To keep matters in context it is desirable to sufficiently to return the favor in kind in a con- Ware and Keith Uncapher describe where this work stood relative to the trolled manner. A survivable command and con- among others at RAND state of the art at the time it took place. In 1964 trol communications infrastructure would be for continuing strong sup- I wrote a highly detailed set of memoranda that mandatory to get away from the guns loaded, port while undertaking describe this work down to the circuit level [1]. hair trigger doctrine of the time. my highly controversial These memoranda are on the Internet and read- RAND computer simulations showed that the activity. ily available to the reader [2]. I devoted an entire telephone system would fail, while most telephone 42 0163-6804/02/$17.00 © 2002 IEEE IEEE Communications Magazine • July 2002 facilities survived. I believed that the problem was obviously in the topology of our communications networks, and there might be a solution. (I had worked on the subject of survivable networks while at Hughes Aircraft before coming to RAND, so I was not new to the subject.) WHY NETWORKS ARE VULNERABLE Let’s consider three different communications network topologies in Fig. 1. (a) The centralized network has all its nodes connected to a central switching node to allow simple switching, giving it a single point of high vulnerability. (b) The decentralized network, representative Link of the AT&T Long Lines network at the time, is better. Instead of a single central switching node, Station the network comprises small centralized clusters, (a) (b) (c) with most traffic going to nearby neighbors, and only the longer distance traffic routed to the I Figure 1. Three forms of networks: a) centralized; b) decentralized; c) dis- longer links. tributed. (c) The distributed network is a network with- out any hierarchical structure; thus, there is no single point of vulnerability to bring down much THE BROADCAST STATION of the network. DISTRIBUTED NETWORK Shortly after I arrived at RAND I began to 2 The optimum degree of study the behavior of distributed networks with My first RAND distributed network proposal in redundancy is a function different levels of redundant connections. 1960 was for a survivable teletypewriter network of the sum of the loss of In Fig. 2, we see a network with its nodes tied to carry what was then called “minimum essen- capacity by component together with the minimum possible number of tial communications.” Carrying briefing charts reliability failures plus the links. This is called a network of redundancy and slides3 around to the Pentagon and various damage anticipated. It level 1. A network of redundancy level 2 looks military command centers, I found the term was doesn’t make much differ- like a fishnet, with horizontal and vertical links. unrealistic. Far, far more capacity was needed ence if the damage is due When we reached redundancy levels on the than was previously realized. So I went back to to an attack or compo- order of 3 an interesting phenomenon occurred: the drawing board and took on the challenge to nent reliability. Thus, it the network became extremely robust. If a node come up with a scalable communications switch- would be possible to build survived physical damage, it would likely be con- ing structure capable of dynamically routing highly reliable networks nected to all other surviving nodes in the largest high-bit-rate traffic among a large set of poten- out of unreliable links and single group of surviving nodes. This meant that tial users, and where user requirements could nodes able to withstand it would be theoretically possible to build extremely not be predicted in advance. high levels of damage at a reliable communication networks out of unreliable (Meanwhile, the broadcast teletypewriter redundancy level of about links, by the proper use of redundancy.2 In other concept crept slowly through the Air Force pro- 3. Even a redundancy words, if a redundantly connected node survived cess and was eventually assigned to the Rome level of 1.5 would more the physical attack, there is a high probability Air Development Center in upper New York than suffice for a network that this node, at least on paper, was somehow State for implementation. An experimental net- where no enemy attack is connected to all the other surviving nodes. work was built to cover the northeast section of anticipated, just compo- “Somehow” was the issue, and was the motiva- the US. Its only stress test was the massive nent reliability. tion for packet switching. Northeast power blackout in 1965 when it was said to have worked well.)4 [4, 5] 3 The majority of the illus- STATE OF THE ART, 1960–1964 trations in this article were NETWORK SYNCHRONIZATION originally projected using In those days (around 1960) we didn’t know how glass lantern slides, to build communication switches where signals My interest was now focused on creating a new, reminding me how much could traverse many serially connected nodes very high-data-rate (in 1960s terms) network. I technology has changed in and operate reliably in the face of damage. The was able to start the design with a clean sheet of 40 years. AT&T telephone system had a limit of five paper rather than fighting all the constraints of switched tandem links before a phone call was our then networks. For example, since the data 4 Extensive testing has unacceptable. flow in the network had to traverse many tandem taken place in the past on A new way was needed to get usable signals nodes, I felt it would be impossible to synchronize the simultaneous trans- through a large number of nodes, traveling via all individual links in tandem to operate at the mission of a commercial highly circuitous paths that could not be deter- exact same data rate. Instead I proposed small broadcast and submodu- mined in advance. The new network would have computer-based switching nodes, to provide a lated teletyped signal.