H ISTORY OF C OMMUNICATIONS

EDITED BY MISCHA SCHWARTZ

AN EARLY HISTORY OF THE INTERNET
LEONARD KLEINROCK

INTRODUCTION
It is impossible to place the origins of the Internet in a single moment of time. One could argue that its roots lie in the earliest communications technologies of centuries and millennia past, or the beginnings of mathematics and logic, or even with the emergence of language itself. For each component of the massive infrastructure we call the Internet, there are technical (and social) precursors that run through our present and our histories. We may see to e!plain, or assume away, whatever range of component technologies we li e. It is e"ually possible to narrow Internet history down to specific technologies with which we are the most familiar. #here are also many individuals that may be said to have $predicted% the

Internet. In &'(), *i ola #esla foresaw +&, a technology that would allow $a business man in *ew -or to dictate instructions, and have them instantly appear in type at his office in .ondon or elsewhere% and would allow global access to $any picture, character, drawing, or print.% #hirty years later, /. 0. Wells articulated +1, his idea of a $World 2rain% as $a depot where nowledge and ideas are received, sorted, summari3ed, digested, clarified and compared.% #hese ideas were followed by a &'45 essay +6, by 7annevar 2ush, predicting a machine with collective memory that he called the meme!, with which $Wholly new forms of encyclopedias will appear, readymade with a mesh of associative trails running through them, ready to be dropped into the meme! and there amplified.% #hese predictions, however, do not

help us understand why the specific events, innovations, people, and circumstances that formed our Internet emerged when they did. 8oing so is not possible from the scale of centuries or single individuals. #his column9s focus is on the defining inventions and decisions that separate early technologies that were clearly not the Internet, from a wide range of recent inventions that may help characteri3e our Internet, but were also built within it. #hus, in this column we trace both the early history of the science and infrastructure that emerged as the :;<:*=#, and the tra>ectory of

development it set for the even broader construct that we now call the Internet. :s one of many individuals who participated in the Internet9s early history, I

also offer a personal account of the same events, as an autobiographical element in this story. In doing so, I aim to further conte!tuali3e publications from the period ? my primary source materials ? with details from firsthand e!perience. #his perspective may add to our depth of historical understanding, in which the e!tent of personal detail does not imply a greater importance to the events presented. In focusing on the wor of individual researchers and developers, I rely on the various publications that followed the wor of these individuals to lin this story to the factual historical record we will follow. #here are, of course, many important personal and institutional stories that have yet to be told. #he @niversity of Aalifornia at .os :ngeles (@A.:) is heavily mentioned in this column, as it was the site of so much foundational wor . I view this period as a synergistic surge of technology, engineered by a magnificent group of researchers and developers amidst a defining period of challenge, creativity, invention, and impact.

other ey contributors and successive

stages of development in Internet history. I present these threads and phases chronologically so we can revisit the history as it unfolded. One may find elaborations on this history in two earlier papers +4, 5,
THE RESEARCH THREAD

BEFORE THE BEGINNING: TWO THREADS THAT MEET
#he Internet did not suddenly appear as the global infrastructure it is today, and neither did it form automatically out of earlier telecommunications. 8uring the late &'5(s and early &'B(s, two independent threads were being woven. One was the research thread that eventually led to the pac et switching networ s of today9s Internet. #his thread followed three possible paths to the technologies that eventually emergedC the researchers involved were, in chronological order, myself, <aul 2aran, and 8onald 8avies. 2elow we e!plore these three paths, which were independently pursued in the "uest to provide data networ ing theory, architecture, and implementation. #he second thread was the creation and growth of the :dvanced ;esearch <ro>ects :gency (:;<:), the institution that funded and deployed these technologies ? a process that, as we will see, was by no means automatic. #hese two threads merged in the mid&'B(s, creating the historical $brea % that led to the :;<:*=#. Once these threads merged, the implementation and deployment phase began, bringing in
26

In Danuary &'5E& I began as a graduate student in electrical engineering at Fassachusetts Institute of #echnology (FI#). It was there that I wor ed with Alaude Ghannon, who inspired me to e!amine behavior as large numbers of elements (nodes, users, data) interactedC this led me to introduce the concept of distributed systems control and to include the study of $large% networ s in my subse"uent thesis proposal. In that FI# environment I was surrounded by many computers and reali3ed that it would soon be necessary for them to communicate with each other. /owever, the e!isting circuit switching technology of telephony was woefully inade"uate for supporting communication among these data sources. #his was a fascinating and important challenge, and one that was relatively une!plored. Go I decided to devote my <h.8. research to solving this problem, and to develop the science and understanding of networ s that could properly support data communications. Aircuit switching is problematic because data communications is bursty, that is, it is typically dominated by short bursts of activity with long periods of inactivity. I reali3ed that any static assignment of networ resources, as is the case with circuit switching, would be e!tremely wasteful of those resources, whereas dynamic assignment (I refer to this as $dynamic resource sharing% or $demand access%) would be highly efficient. #his was an essential observation, and in &'5' it launched my research thread as I sought to design a new ind of networ . Its architecture would use dynamic resource allocation to support the bursty nature of data communications, and eventually provide a structure for today9s pac et-switched networ s.
1 Later that year on October 4, I experienced a

widely shared feeling of surprise and embarrassment when the Soviet Union launched Sputni , the first artificial !arth satellite" In response, #resident !isenhower created $%#$ on &ebruary ', 1()* to regain and maintain U"S" technological leadership"

(&B6-B)(4H&(HI15.(( J 1(&( I===

I=== Aommunications Faga3ine K :ugust 1(&(

H ISTORY OF C OMMUNICATIONS
#his concept of resource sharing was emerging at that time in a totally different conte!tL that of timesharing of computer power. #imesharing was based on the same fundamental recognition that users generate bursty demands, and thus e!pensive computer resources were wasted when a computer was dedicated to a single user. #o overcome this inefficiency, timesharing allocated the computer to multiple users simultaneously, recogni3ing that while one user was idle, others would li ely be busy. #his was an e!"uisite use of resource sharing. #hese ideas had roots in systems li e G:0= +B, and in the FI# Aompatible #ime-Gharing Gystem (A#GG +E,), developed in &'B& by Fernando Aorbato (among the first timesharing systems to be implemented). #he principles and advantages of timesharing were ey to my reali3ation that resource sharing of communication lin s in networ s could provide for efficient data communications, much li e the resource sharing of processors in timeshared systems was accomplishing. In addition, there was already an e!ample of a special-purpose data networ that used resource sharingL the store-and-forward telegraph networ . #he challenge I faced was to create an appropriate model of general-purpose data communications networ s, to solve for their behavior, and to develop an effective design methodology for such networ s. #o do this, I sought to develop a model with dynamic resource sharing, incorporating the fact that data traffic was unpredictable as well as bursty. In order to clear up some misconceptions regarding what I and other investigators were doing in the field in the early days, I will devote some space in the following paragraphs to discuss the relationship between dynamic resource sharing and pac et switching, where the latter is but one of many ways to reali3e the former. #he basic structure I chose was that of a "ueue since it is a perfect resource sharing mechanism. : "ueue is dynamic, adaptive, and efficient, and does not wait for a message that is not there, but rather transmits a message already waiting in the "ueue. Foreover, the performance measures one considers in "ueueing theory are response time, throughput, efficiency, buffering, priorities, and so on, and these are >ust the "uantities of interest in data networ s. In the late &'5(s, the published literature contained almost no wor on networ s of "ueues. /owever, a singular e!ception to this was the wor by Dames Dac son, who published a classic paper +), on open networ s of "ueues. :s we see below, I was able to apply Dac son9s result to represent the data networ s of interest by ma ing serious modifications to his model. Go the stage was setL #here was a need to understand and design generalpurpose data communication networ s that could handle bursty data traffic, there was an emerging approach based on resource sharing in timeshared systems, there was an e!isting special-purpose networ that suggested it could be done, and there was a body of "ueueing theory that loo ed promising. :s a result, I prepared and submitted my FI# <h.8. thesis proposal +', in Fay &'B&, entitled $Information Flow in .arge Aommunication *ets% in which I developed the first analysis of data networ s. I chose a "ueueing theoretic model based on Dac son9s model to characteri3e a data networ as a networ of communication channels whose purpose was to move data messages from their origin to their destination in a hop-byhop fashion. =ach channel was modeled as a resource serving a "ueue of data messages awaiting transmissionC I discussed how $#he nets under consideration consist of nodes, connected to each other by lin s. #he nodes receive, sort, store, and transmit messages that enter and leave via the lin s....% Fy underlying model assumed that the stream of messages had randomly chosen lengths and, when applied to data networ s, yielded a problem whose e!act solution turned out to be hopelessly intractable. I altered the model and also introduced a critical mathematical assumption, the Independence :ssumption,1 which tamed the problem and allowed for an elegant solution. With this solution, I was able to solve for the many performance measures of these networ s. For e!ample, I showed that by scaling up the networ traffic and bandwidth properly, one could reduce the system response time, increase the networ efficiency, and increase the networ throughput, all simultaneously +&(,. In the course of e!amining data networ performance, it became clear to me that it was important to e!plore the manner in which mean response time was affected when one introduced a priority "ueueing discipline on the traffic. I chose to understand this influence in the case of a single node first and then to apply the results to the general networ case. #his led to a publication in :pril
,hapter - of my dissertation .1/0 elucidates this problem and the role of the Independence
+

$ssumption"

&'B1, which turned out to be the first paper +&&, to introduce the concept of brea ing messages into smaller fi!edlength pieces (subse"uently named $pac ets,% as e!plained below). In it I provided a mathematically e!act analysis of the mean response time, and showed the advantages to be gained by utili3ing pac et switching for this new networ .6 *ote that the fi!ed length pac ets I introduced did not match the randomly chosen lengths of the model, but fortunately, the ey performance measure I solved for, the overall mean system response time, did not re"uire that assumption, so the mathematical model properly reflected the behavior of fi!ed length pac ets as well. I also developed optimal design procedures for determining the networ capacity assignment, the topology, and the routing procedure. I introduced and evaluated distributed adaptive routing control procedures, noting that networ Hrouting control is best handled by sharing control among all the nodes rather than relegating control to one or a small number of nodes. #his distributes the control load (thereby not unduly loading any one node), introduces the ability to change routes on the fly dynamically (based on current load, connectivity, and destination address), enables the networ

to scale to a very large number of nodes, and dramatically improves the robustness of the networ . Whereas my focus was not principally on the engineering details of pac et networ s, I did address engineering details when I built a complete networ simulation model and conducted e!tensive simulation e!periments confirming the correctness of the theory. #hese e!periments included detailed message bloc s (with headers, origin and destination addresses, priority indicators, routing labels, etc), dynamic adaptive routing tables, priority "ueueing structures, traffic specifications, and more.4 <ac eti3ation was an integral part of a much broader body of nowledge that had to be developed to prove the case for data networ s. Indeed, pac eti3ation alone was not the underlying technology
- One of the

important advantages of using pac 1 ets turned out to be that short messages would not get 2trapped3 behind long messages4 I was able to show this gain in response time exactly"
4 $ccess

to my simulation notes can be found at http566ucla"worldcat"org6title6leonard1 lein1 roc scorrespondence1course1and1research1 notes11(/11 1('+6oclc6+/-1/4(/47referer8brief9results

I=== Aommunications Faga3ine K :ugust 1(&(

27

HISTORY OF COMMUNICATIONS
that led to :;<:*=# design fundamentals. #o be sure, pac eti3ation was and remains a core element of today9s networ ing technology, but it is not identical to networ efficiency. ;ather, the fundamental gain lies in dynamic resource sharing. It is important to point out that there are many ways in which dynamic resource sharing can be accomplished, with pac et switching being only one such methodC other methods include polling +&1,, message switching +&6,, asynchronous time-division multiple access (:#8F:) +&4,, carrier sense multiple access with collision detection (AGF:HA8) +&5,, and others. I completed and filed my <h.8. dissertation +&B, in 8ecember &'B1, having created a mathematic al theory of pac et switching for dynamic resource sharing, thus providing the fundamenta l underpinnin gs for :;<:*=# technology. I showed that these networ s were efficient, stable, scalable, robust, adaptive, and, most of all, feasible. 8ecades of important research on these topics have since ta en place around the world. 2y the time my dissertation was published as the first boo +&E, on computer networ s in &'B4, the idea of pac eti3atio n itself was appearing more broadly. #he ne!t contributor to pac et switching was <aul 2aran of the ;:*8 Aorporation, who was busy wor ing on military command and control systems during the early &'B(s with the goal of using redundancy and digital technology to design a robust multilateral military communicati ons networ . /e recogni3ed the vulnerability of the telephone networ due to its centrali3ed architecture. In Geptember &'B1 he published a paper +&), on how $hot potato% adaptive alternate routing procedures and distributed principles could utili3e a $standard message bloc ,% also to fall under the $pac et% umbrella, which will be addressed below. /is purpose was to create a networ capable of functioning after a Goviet nuclear attac +&',. In :ugust &'B4 he produced

a set of && important reports +1(, reinforcing his prior description with simulations and elaborating on many details of the design. /e, too, discovered the importance of going to digital networ s and of the robustness provided by distributed routing. /e attempted to get :#M# to implement the design, but failed to convince them (presumably due to their analog mindset). In &'B5 ;:*8 approached the :ir Force to implement it, but they deferred to the 8A:5C at this point, 2aran decided not to pursue the implementati on any further. 2aran9s

wor was done independentl y of the wor that I had done earlier at FI# and, in many ways, the results we achieved in addressing the problem of pac et networ s were complement ary. #he third early contributor to pac et switching was 8onald 8avies, of the *ational <hysical .aboratory (*<.) in the @nited Ningdom. /e began thin ing about pac et networ s in &'B5 and coined the term $pac et% that year. In a privately circulated paper +1&, dated Dune &'BB, he described his design for a data networ and used my earlier theory to calculate its performance . 8avies lectured to a public audience in Farch &'BE, recommendi ng the use of his technology for the design of a public switched data networ , and published an October &'BE paper +11, with his *<. group

in which details of the design were first described in an open publication. #his plan was for an *<. 8ata Aommunicati ons *etwor , but the @.N. 8epartment of #rade and Industry only authori3ed the implementati on of one node. #hat node became operational in &'E(. Further details of a full networ design were described by the *<. team in &'B) +16, 14, and &'B' +15,C it is not clear where a multiplenode deployment by this team might have led, but it obviously had potential. #his reluctance to support an *<. pac etswitched networ was reminiscent of the view ta en by :#M# and 8A: in not supporting an implementation of the ;:*8 wor . #he wor of 2aran and 8avies focused on the engineering and architectural issues of the networ design. Fy wor emphasi3ed and provided the mathematical underpinning s and supporting simulation

e!periments of the networ analysis and design, including optimi3ation as well as formulating the basic principles of pac et networ s that include dynamic resource sharingC this "uantitatively showed that these networ s were feasible. Fy tra>ectory was more fortunate as the :;<: thread rolled out and adopted my principles for their design of the :;<:*=#, and provided me the opportunity to participate in its implementati on and deployment. 8ifferent tra>ectories were ta en by 2aran and then later by 8avies, with 2aran9s unsuccessful attempts to get his ideas implemented and with 8avies9 frustration by the footdragging of the @.N. government. It was not enough to
:efense ,ommunicati ons $gency, which was renamed in 1((1 to be today;s <+=1=> :efense Information Systems $gency ? :IS$"
)

put good ideas forward, but it was also necessary to prove that the concepts were "uantitativel y sound, and then to implement and deploy an operational networ that would bring these ideas and designs to use.
THE ARPA THREAD

.et us step bac chronologic ally and now pursue the second threadL the role of :;<: in defining the need for a data networ , putting the management structure in place to enable its development, and providing the funding necessary for its implementat ion and deployment. D. A. ;. .ic lider ($.ic %) entered the story when he published his landmar &'B( paper +1B, $FanAomputer Gymbiosis.% /e defined the title as $an e!pected development in cooperative interaction between men and electronic computers.% #his wor

envisaged a system $to enable men and computers to cooperate in ma ing decisions and controlling comple! situations without infle!ible dependence on predetermined programs%C he had seen such a fle!ible system in the aforemention ed G:0= system. Once again, we find a forecast of what future telecommuni cations might provide ? and .ic was perhaps the first to write at a time when viable ways to create that future were emerging. :lthough a visionary, .ic was not a networ ing technologist, so the challenge was to finally implement such ideas. In Fay &'B1 .ic and Welden Alar outlined their views on how networ ing computers could support social interaction, and provide networ ed access to programs and data +1E,. #his e!tended his earlier ideas of what he now referred to as a 0alactic *etwor (in fact, he

nic named his group of computer e!perts $#he Intergalactic *etwor %). .ic was appointed as the first director of :;<:9s newly formed Information <rocessing #echni"ues Office (I<#O) in October &'B1. /e "uic ly funded new research into
28

advanced computer and networ ing technologies as well as areas that involved man-computer interaction and distributed systems. 2y the end of &'B1, .ic had articulated his grand vision for the 0alactic *etwor , of which I was

unaware, and I had laid out the mathematical theory of pac et networ s, of which .ic was also unaware. #hese ideas would soon intersect and reinforce each other in a series of ey events between &'B1 and &'B'. I >oined the @A.: faculty in &'B6. .ic

I=== Aommunications Faga3ine K :ugust 1(&(

H ISTORY OF C OMMUNICATIONS
passed the directorship of I<#O to Ivan Gutherland, an FI# colleague of mine, in Geptember &'B4. In that role Gutherland wished to connect @A.:9s three I2F mainframes in a three-node on-campus computer networ , which would have been easy to accomplish with the means I had laid out in my <h.8. dissertation. /owever, the @A.: networ was never reali3ed due to administrative discord. *evertheless, the seeds for an :;<:funded networ had now been sown. =arly the ne!t year (&'B5), Gutherland awarded .arry ;oberts (another FI# colleague of mine who was "uite familiar with my networ ing research) a contract to create a dialup &1(( bHs data connection across the @nited Gtates. .ater that year, ;oberts accomplished this in collaboration with #homas Farill, demonstrating that such a connection re"uired a different, more sophisticated networ than the telephone networ offered +1),. Feanwhile, at :;<:, Gutherland recruited ;obert #aylor to become associate director of I<#O in &'B5. While there, #aylor also recogni3ed the need for a networ , this time specifically to connect :;<: research investigators to the few large e!pensive research computers across the country. #his would allow them to share each other9s hardware, software, and applications in a cost-effective fashion.B #aylor then dropped into the office of the :;<: director, Aharlie /er3feld, to re"uest funding for this nascent networ ing pro>ect. /er3feld was a man of action who new how to ma e a fast decision, and within 1( minutes he allocated I& million to #aylor as initial funding for the pro>ect. #aylor, who had since suc/ @his sharing of resources was the primary

instituted beside the original research1driven networ >"

ceeded Gutherland as I<#O director in :ugust &'BB, brought in ;oberts as the I<#O chief scientist that 8ecember. 2ringing ;oberts in to manage the networ ing pro>ect turned out to be a critical hire as ;oberts was to contribute at all levels to the coming success of data networ ing. #he research and :;<: threads had now merged, and the pro>ect would soon become the :;<:*=#. #hese were critical steps in Internet history, for not even in the post-war @nited Gtates did technological progress flow directly from ideas. In contrast to refusals from the private sector to fund the beginnings of the pro>ect, :;<: made available the will and funding of the @.G. government.E :;<:9s management and support fostered the early culture of shared, open research that was crucial to the success of the :;<:*=# program.

THE BEGINNING: THE ARPANET LAUNCH
#he commitment to create the :;<:*=# was now in play. ;oberts was empowered to develop the networ concept based on .ic 9s vision, my theory, and #aylor9s application. #here were basically two matters to be considered in this pro>ect. One was the issue of creating the switches and lin s underlying the networ infrastructure, with the proper performance characteristics, including throughput, response time, buffering, loss, efficiency, scalability, topology, channel capacity, routing procedure, "ueueing discipline, reliability, robustness, and cost. #he other was to create the appropriate protocols to be used by the attached (host) computers) so that they could properly communicate with each other. Ghortly after his arrival, ;oberts called a meeting of the :;<: <rincipal Investigators (<Is) in :pril &'BE at the
' In sharp contrast to $%#$;s enthusiasm for

motivation for creating the $%#$A!@" #aul Baran developed a networ design <described above> that would maintain communications ? and specifically, Second Stri e ,apability ? in the event of a nuclear attac by the USS%" Cis and my wor served different aims" Dhen $%#$ began wor on the $%#$A!@, my wor was used for the reasons described herein" Cis application to military communications gave rise to the myth that the $%#$A!@ was created to protect the United States in case of a nuclear attac " @his is not to ta e away from Baran;s accomplishments4 indeed, by the time the $%#$A!@ began in 1(/(, he had moved on to different proEects, including the Institute for the &uture <he stepped bac into the $%#$ foray in 1('4F1(') to recommend that an early commercial version of the $%#$A!@ be

networ ing, in the early 1(/=s, when I intro1 duced the ideas of pac et1switched networ s to what was then the world;s largest networ ing company, $@7@, I met with narrow1 minded and failed thin ing, and was summarily dismissed by them" @hey commented that pac et switching would not wor , and even if it did, they wanted nothing to do with it" Baran had a similar reaction from $@7@"

*$

maEor challenge for such a networ was that it would connect computers with incompatible hardware and software"

@niversity of Fichigan, where :;<:*=# planning was discussed in detail. It was there that the basic specifications for the underlying networ were debated among us <Is. For e!ample, Wesley Alar put forward the concept of using an unmanned minicomputer at each location to handle all of the switching and communications functionsC it was to be called an Interface Fessage <rocessor (IF<). #his would offload the networ ing functions from the host, greatly simplify the design by re"uiring only one interface to be written for each host to the standard IF<, and at the same time would decouple the networ design from any specific host hardware and software. :nother specification had to do with the measure of reliability of the planned networ C this we specified by re"uiring that the topological design ' produce a $two-connected net,% thus guaranteeing that no single failure would cause any non-failed portion of the networ to lose connectivity. -et another re"uirement we introduced was for the networ to provide an e!perience as if one were connected to a local timeshared computer even if that computer was sitting thousands of miles across the networ C for this we specified that short messages should have response times no greater than 5(( ms

(the networ design provided 1(( ms at its inception). Foreover, since this was to start out as an e!perimental networ , I insisted that appropriate measurement tools be included in the IF< software to allow for tracing of pac ets as they passed across the networ , ta ing of snapshots of the IF< and host status at any time, artificial traffic generation, gathering and forwarding of statistics about the networ , and a mechanism for controlling these measurements. Following this :pril meeting, ;oberts put together his outstanding plan for the :;<:*=# design and presented it as a paper +1', at a conference in 0atlinburg, #ennessee in October &'BE. :t this conference, ;oger Gcantlebury of the *<. also presented their aforementioned >ointly published paper +11, describing a local networ they were developing. It was during a conversation with Gcantlebury at this meeting that ;oberts first learned of the *<. wor as well as some details of the wor by 2aran at ;:*8. #he research by myself at FI#, by 2aran at ;:*8,
@o assist with the topological design, Aetwor $nalysis ,orporation <A$,>, whose ,!O was Coward &ran , was brought in as a contractor"
(

I=== Aommunications Faga3ine K :ugust 1(&(

29

H ISTORY OF C OMMUNICATIONS

UC Berkeley Stanford U. SRI UC Santa Bar ara U. Uta# R AND UC$A AR&A net'ork to(olo)y *e+a,(leSDC Wa"#!n)ton Un!.. &ro/. MAC &enta)on U. M!%#!)an Carne)!e Mellon U. B.T.$. BB N Har.ard U. Dart,o0t# Colle)e

U. Ill!no!" $!n%oln $a

Figu ! "# 1(1node $%#$A!@ as shown in the original %&G"

and by 8avies, Gcantlebury, et al" at *<. had all proceeded independently, mostly without the researchers nowing about the others9 wor . #here was, though, some cross-fertili3ationL 8avies had used my analytical model for data networ s in his wor C as a result of discussions at this conference, ;oberts adopted 8avies9 word $pac et% for the small fi!ed length pieces I had suggested we brea messages into, and which 2aran referred to as $message bloc s%C its fi!ed length was chosen to be &(14 bits for the :;<:*=# design (both 2aran and 8avies had suggested this same length)C as a result of the discussion with Gcantlebury, ;oberts decided +6(, to upgrade the bac bone line speed from '.B bHs to 5( bHs for the :;<:*=# design. Following these &'BE meetings, a se"uence of drafts for the IF< specification was prepared.&( #his culminated in Farch &'B) when ;oberts and 2arry Wessler produced the final version of the IF< specification, which they then discussed at an :;<: <I meeting later that month. On Dune 6, &'B), the :;<:*=# <rogram <lan +6&, was formally submitted to :;<: by ;oberts, and it was approved on Dune 1&, &'B). #he :;<:*=# procurement process was now officially underway. 2y the end of Duly &'B), a ;e"uest
1= $mong those involved in these first drafts

were &ran Destervelt, !lmer Shapiro, Hlen ,uller, and myself" $%

for Ouotation (;FO) +61, for the networ IF<s was mailed to &4( potential bidders. #he &'-node e!ample to be delivered by the contractor is shown in Fig. &. #he handling of data streams specified that the hosts would communicate with other hosts by sending messages (of ma!imum length )&'1 bits) to their attached IF<s, that these messages would be bro en into pac ets (of ma!imum length &(14 bits each ? thus, at most ) pac ets per message) by the IF<, and that IF<s would communicate with each other using these pac ets. #he movement of pac ets through the subnetwor of IF<s was to be controlled by a distributed dynamically updated routing algorithm based on networ connectivity and loading as well as pac et destination and priority. =rrors in pac et transmission between IF<s were managed by error detection and retransmission. <ac ets were to be reassembled into their original messages at the destination IF< before delivery to the destination host. #he basic structure of this IF< specification contained contributions from a number of individuals, including my own research. ;oberts had been well aware of my wor since my time at FI#, where we were officemates, later stating,&& $In order to plan to spend millions of dollars and sta e my reputation, I needed to understand that it would wor . Without Nleinroc 9s wor of *etwor s and Oueueing #heory, I could never have ta en such a radical step.% +66,

#he ;FO resulted in &1 proposals being submitted in :ugust &'B) (notably missing were I2F and :#M#). :s these proposals were being evaluated at :;<:, ;oberts awarded a research contract to me at @A.: in October to create the *etwor Feasurement Aenter (*FA). #he tas of the *FA was to measure the behavior of the :;<:*=# by conducting e!periments to determine its faults, performance, and outer limits (through the use of stress tests). I was fortunate to have a star team&1 of graduate student researchers, developers, and staff for this pro>ectC a number of these appear in continued roles later in this story. : wee before Ahristmas &'B), 2olt, 2erane and *ewman (22*) won the competitive bid and was awarded the contract to develop the IF<-to-IF< subnetwor . #he 22* team,&6 supervised by Fran /eart, produced some remar able accomplishments. #his team had selected the /oneywell 88<-5&B minicomputer with &1 b of memory for the program to be the machine on which
11 %oberts also goes on to say that my

dissertation was 2critical to my standing up to them and betting it would wor "3
1+ Iey

members of my U,L$ team included a research team <Jerry ,ole, $l :obies i, Hary &ultK, Lario Herla, ,arl Csu, Jac Meigler>, a software team <Nint ,erf, Steve ,roc er, Herard :eLoche, ,harley Iline, Bill Aaylor, Jon #ostel>, a hardware engineer <Li e Dingfield>, and others"

I=== Aommunications Faga3ine K :ugust 1(&(

H ISTORY OF C OMMUNICATIONS
he was most li ely the first programmer on the Internet by virtue of having done code design for the IF< in their &'B) response to the ;FO. Whereas members of the 22* team were busy testing the IF<9s ability to provide IF<to-IF< data e!changes, testing the behavior of a networ of IF<s was difficult to do in a laboratory environmentC the true behavior was more properly tested in the deployed networ with real traffic and with many nodes, which is e!actly what the *FA was designed to do. 2asically, 22* was given less than nine months to deliver the first IF< to
1- Iey members of Ceart;s team included Ben

SRI U TA H UC$A

1DS 234 IM& I M &

Bar er, Bernie ,osell, Dill ,rowther, %obert Iahn, Severo Ornstein, @ruett @hach, :ave Dalden, and others"

IM& 564789 IBM I M & UCSB

1DS 1:8

@A.: by early Geptember &'B'. #heir performance was outstanding. #he first IF< at @A.: was to be followed by the second IF< in October to G;I, the third IF< in *ovember to the @niversity of Aalifornia at Ganta 2arbara (@AG2), and the fourth IF< in 8ecember to the @niversity of @tah. #he initial networ was to be that shown in Fig. 1. #hese four sites were selected due to their ability to provide speciali3ed networ services andHor support. Gpecifically, @A.: (connecting an G8G Gigma-E /ost computer) would provide the *FA (under my supervision), G;I (connecting an G8G '4( host computer) would provide 8oug =nglebart9s /uman Intellect :ugmentation Gystem (with an early version of hyperte!t in his *.G system) as well as serve as the *etwor Information Aenter (under =li3abeth +Da e, Feinler9s supervision), @AG2 (connecting an I2F 6B(HE5 host computer) would provide interactive graphics (under 0len Auller9s and 2urton Fried9s supervision), and the @niversity of @tah (connecting a 8=A <8<-&( host computer) would provide advanced 68 graphics (under the supervision of Ivan Gutherland). #he fact that /eart and his team at 22* succeeded in delivering this new technology with new applications and new users in an ontime, on-budget fashion was incredible. 2ut this contract to develop the underlying networ was only the first of the two ey tas s that were needed to

DEC &D&:;4

Figu ! 2# @he initial four1node $%#$A!@ <1(/(>"

the IF< would be basedC they were contracted to implement the IF< functions by modifying the hardware and software of the 88<-5&B, to connect these IF<s to long-haul 5( bHs lines leased by ;oberts from :#M# under the 8o8 #elpa tariff, and to deploy the subnetwor . #he 22* team developed an elegant host-IF< design that met the :;<: specificationsC this specification was written as 22* ;eport &)11 +64, by ;obert Nahn, who was in charge of the system design at 22* (Nahn appears later in this story in some very significant roles, as we shall see below). One of the 22* team, 8ave Walden, points out that

deploy a wor ing pac et-switched networ . ;ecall that the other tas was to create the appropriate protocols to be used by the attached (host) computers so that they could properly communicate with each other. #his second tas was assigned to the four chosen :;<:*=# research sites to figure out on their own. #hus began another thread of innovative development that characteri3ed the :;<:*=# culture. #his thread actually begins in the summer of &'B) when =lmer Ghapiro of G;I, in response to a re"uest by :;<:, called a meeting of programmers from among those first sites that were to be connected into the :;<:*=#. #heir main charge was to study and resolve the issues of host-tohost communication. <resent at this meeting was one programmer from each of the first four sites to receive IF<s as followsL Gteve
I=== Aommunications Faga3ine K :ugust 1(&(

Aroc er (@A.:), Deff ;ulifson (G;I), ;on Gtoughton (@AG2), and Gteve Aarr (@niversity of @tah). #his group, plus the many others who >oined later, were soon to be named the *etwor Wor ing 0roup (*W0) with Ghapiro its first chairman.&4 @A.:9s Don <ostel served as the ;e"uest for Aomments (;FA) editor (a role he held until his untimely death in &'')). #hey had no official charter against which to wor , and so were afforded the uni"ue opportunity to invent and create as needed. #here was no sense of "ualifying membershipC all one had to do was to contribute and participate. #heir focus moved to the creation of high level interactions and, eventually, to the notion of a layered set of protocols (transport services below a set of application-specific protocols). 2asically, this was a highly resourceful, self-formed, collegial, loosely

configured group of maveric graduate students who we (the :;<: <Is) had empowered to design and implement the protocols and software for the emerging networ . #hey too on the challenge we ceded to them and
14 @he

names of some of the other ey individuals who participated early on in the ADH include Bob Braden, Nint ,erf, :anny ,ohen, Bill :uvall, Lichel !lie, Jac &einler, Jon #ostel, and Joyce %eynolds"
1) It

is remar able how effective the %&,s, the ADH and the I!@& have served the networ community" In spite of the fact that they are loosely structured and involve large numbers of outspo en professionals, they have been able to move forward on a number of critical Internet issues" $"

HISTORY OF COMMUNICATIONS

Figu ! $# @he entry
in the original IL# log, which is the only record of the first message transmission on the Internet"

created an enduring *W0 structure that later led to today9s Internet =ngineering #as Force (I=#F).&5 Once the IF<-host specification was released by 22* in the spring of &'B', the *W0 began to focus on the lowerlevel issues such as message formats. #hey decided to e!change ideas through a very informal set of notes they referred to

as $;e"uests for Aomments% (;FA). #he first ;FA +65,, entitled $/ost <rotocol,% was written by Aroc er in :pril &'B'. Aroc er became the second Ahairman of the *W0 early on. We now had the two main :;<:*=# development efforts underwayL P : formal contract with 22* to

create the IF<IF< subnetw or P :n informal group of program mers (mostly graduate students) who were charged with developi ng the /ost-to/ost <rotocol #hings began to move rapidly at this point. #he date of the first IF< delivery, scheduled to arrive to us at @A.: in

early Geptember &'B', was fast approaching. Feanwhile, at the *FA, we were busy collecting data so that we could predict performance of the networ based on my earlier theory. For this, it was necessary to estimate the traffic loads that the host sites would present to the networ . ;oberts and I contacted a number of the early sites and as ed them how much traffic they e!pected to generate and to which other sites. We also as ed them how much traffic they would allow into their sitesC to my surprise, many refused to allow any traffic from the networ to use their

hosts. #heir argument was that their hosts were already fully utili3ed serving their local customer base. =ventually they relented and provided their e!pected traffic loads. #hat traffic matri! was used in the Duly &'B) ;FO +61, and in a paper I published +6B,, thereby sealing their commitment . On Duly 6, &'B', two months before the IF< was due to arrive, @A.: put out a press release +6E, announcing the imminent deployment of the :;<:*=#. In that release I described what the networ would loo li e, and what would be a typical application. I am "uoted in the final paragraph as saying, $:s of now, computer networ s are still in their infancy, but as they grow up and become more sophisticate d, we will probably see the spread of Qcomputer utilities,9 which, li e

present electric and telephone utilities, will service individual homes and offices across the country.% It is gratifying to see that the $computer utilities% comment anticipated the emergence of web-based I< services, that the $electric and telephone utilities% comment anticipated the ability to plug in anywhere to an always on and $invisible% networ , and that the $individual homes and offices% comment anticipated ubi"uitous access. /owever, I did not foresee the powerful social networ ing side of the Internet and its rapidly growing impact on our society. On Gaturday, :ugust 6(, &'B', the first IF< arrived at @A.:. On Geptember 1, the day after .abor 8ay, it was connected via a &5-foot cable to the @A.: host computer, our G8G Gigma-E machine. #his established the first

node of the fledgling networ , as bits moved between the IF< and the Gigma-E. #his is often regarded as a very significant moment in the Internet9s history. In early October the second IF< was delivered by 22* to G;I in Fenlo <ar , Aalifornia. #he first high-speed lin of what was to become the Internet was connected between those two IF<s at the $bla3ing% speed of 5( bHs. .ater in October, G;I connected their G8G '4( host computer to their IF<. #he :;<:*=#9s first host-tohost message
$2

was sent at &(L6( p.m. on October 1', &'B' when one of my programmers, Aharley Nline, and I proceeded to $login% to the G;I host from the @A.: host. #he procedure was for us to type in $log,% and the system at G;I was set up to be clever enough to fill out the rest of the command, adding $in,% thus creating the word $login.% Aharley at our end and 2ill 8uvall at the G;I end each had a telephone headset so they could communicat e by voice as the message

was being transmitted. :t the @A.: end, we typed in the $l% and as ed G;I $did you get the lR%C $got the l% came the voice reply. We typed in the $o,% $did you get the oR,% and received $got the o.% @A.: then typed in the $g,% as ed $did you get the gR,% at which point the system crashedS #his was "uite a beginning. Go the very first message on the Internet was the prescient word $lo% (as in, $lo and beholdS%). #his, too, is regarded as a very significant moment in the Internet9s history. #he only record of this event is an

I=== Aommunications Faga3ine K :ugust 1(&(

HISTORY OF COMMUNICATIONS
entry in our IF< log recording it as shown in Figure 6. /ere we see that on October 1', &'B', at &(L6( pm, we at @A.: $#al ed to G;I /ost to /ost.% In *ovember and 8ecember the IF<s and hosts at @AG2 and the @niversity of @tah were connected, respectively, thus completing the initial fournode networ . Further IF< deliveries were halted until we had an opportunity to test this four-node networ , and test it we did. :mong other things, we were able to confirm with measurements some of our theoretical models of networ delay and throughput as presented by 0erry Aole +6),. #he :;<:*=# had now been launched. We now turn to the story of its rollout through its first decade. widely circulated in the community and spread information of the then-current technology that had been deployed. (#wo years later, in Fay &'E1, another ey session at the same conference was devoted to the presentation of five papers +4&, that updated the :;<:*=# state of the artC this, too, was pac aged into a second special :;<: pamphlet.) In mid-&'E( the first cross-country lin was added with a connection from @A.: to 22*, and by Duly the networ contained &( IF<s. #he net grew to &5 IF<s by Farch &'E&. In Geptember &'E& 22* introduced a terminal interface processor (#I<) that conveniently would allow a terminal to connect directly to the :;<:*=# without the need to connect through an attached host. .ater in the year, 22* slipped in a $minor% feature called electronic mail. =lectronic mail had e!isted since the mid-&'B(s for standalone timeshared computer systems, but in late &'E& at 22*, ;ay #omlinson added a small patch to it that allowed the mail to pass between different computers attached to the :;<:*=# using an e!perimental filesharing networ program called A<-*=#. Once he saw that it wor ed, he sent an email message to his group at 22* announcing this new capability, and so $#he first use of networ email announced its own e!istence.% +41,. #his capability went out as a general #=*=T release in early &'E1. 2y Duly &'E1, ;oberts added a management utility to networ email that allowed listing, selective reading, filing, forwarding, and replying to email messages. In less than a year email accounted for the ma>ority of the networ traffic. #he networ 9s ability to e!tend communication between people was becoming evident, a nascent image of .ic 9s vision. .ater that year, in October &'E1, the first public demonstration of the :;<:*=# technology too place at the International Aonference on Aomputer Aommunications (IAAA) in Washington, 8A. Nahn, who by now had been hired into :;<: by ;oberts, organi3ed this large and very successful demonstration in which do3ens of terminals in Washington accessed do3ens of host computers throughout the @nited Gtates in a continuously reliable fashion for the three-day duration of the conference. #he reaction of the computer manufacturers to this :;<:*=# phenomenon was to create proprietary networ architectures based on their own brand of computers.&B #he tele$mong the proprietary networ s were IBL;s SA$ and :!,;s :!,net"
1/

THE FIRST DECADE: FOUR NODES AND THEN THE WORLD
2y the time the first four nodes were deployed in 8ecember &'B', ;oberts (who had succeeded #aylor in Geptember to become the I<#O director) once again met with the *W0 and urged them to e!tend their reach beyond what they had articulated in their first ;FA +65,, $/ost <rotocol.% #his led them to develop a symmetric /ost-to-/ost <rotocol, the first implementation of which was called the *etwor Aontrol <rogram (*A<) and was described by Aroc er in ;FA 6B in Farch &'E( +6',. #his protocol stac was to reside in the host machines themselves and included a hierarchy of layered protocols to implement more comple! protocols. :s *A< began deployment, the networ users could begin to develop applications. #he *A< was the first protocol stac to run on the :;<:*=#, later to be succeeded by #A<HI<. #he tra>ectory of protocol stac development touched on below is another e!ample of multiple possible paths that led the way from the :;<:*=# as it evolved into the Internet. :fter the short evaluation period following the initial four-node deployment, a continual succession of IF<s and networ s were then added to the :;<:*=#. In Fay &'E(, at the :FI<G Gpring Doint Aomputer Aonference, a landmar session was devoted to the presentation of five papers +4(, regarding the newly emerging :;<:*=# technologyC these papers were pac aged into a special :;<: pamphlet that was

phone company continued to ignore it, but the open networ that was the :;<:*=# thrived. Goon, additional networ s were added to the :;<:*=#, the earliest of which were those whose origins came out of wor on wireless networ ing. Aonnecting the :;<:*=# with these different networ s proved to be a feasible but not seamless interoperability issue, and it received a great deal of attention. #he interconnection of networ s was referred to as $internetwor ing% during the &'E(s, a neologism from which the e!panded :;<:*=# was eventually renamed as the Internet. .et us briefly trace the wor on wireless networ ing that led to these additional networ s, which themselves forced attention on improving interoperability solutions. :s pointed out above, these networ s were based on
I=== Aommunications Faga3ine K :ugust 1(&(

wireless multi-access communications in which a shared channel is accessed by many users. 2y late &'E(, *orm :bramson had developed :loha*et +46, in /awaii, a 'B(( bHs pac et radio net based on the novel $unslotted (pure) :.O/:% multi-access techni"ue of random access. In this scheme (unsynchroni3ed) terminals transmit their fi!ed length pac ets at any time over a shared channel at random timesC if more than one transmission overlaps (i.e., collides), then destructive interference prevents any of the involved pac ets from succeeding. #his tolerance of collisions was a departure from the more standard methods of wireline communications to control multi-access systems that used demand access methods ("ueueing, polling, etc., as mentioned earlier) and allowed only one transmission at a time (thus precluding

such collisions). In &'E6 :bramson calculated the capacity of the unslotted :.O/: system +44,, which had a ma!imum efficiency of &) percent, and in &'E1 ;oberts calculated the capacity of a synchroni3ed version (i.e., slotted :.O/:) +45, whose capacity was doubled to 6E percent. /owever, these analyses ignored an essential issue with random access to shared channelsL that they are fundamentally unstable, and some form of dynamic control was needed to stabili3e them, for e!ample, a bac off algorithm to control the way in which collided transmissions are retransmitted. #his stability issue was first identified and addressed by .am and myself +4B, 4E,. It is interesting to note that the :.O/: systems studies eventually led to an investigation of carrier sense multiple access (AGF:) as another wire$$

HISTORY OF COMMUNICATIONS
less access method. AGF: itself led ;obert Fetcalfe to consider a variation called AGF: with collision detection (AGF:HA8), which was the basis for the original =thernet development. 2ased on these concepts, Fetcalfe and 8avid 2oggs implemented AGF:HA8 on a coa!ial cable networ , which was up and running by *ovember &'E6. In sum, they created the =thernet, which is today perhaps the world9s most pervasive networ ing technology +4),. =thernet is crucial to the story of *A< and #A<HI<, for researchers at Tero! <:;A built on this technology in efforts to address the challenges of internetwor ing. Implemented in &'E4 and published in &'E5, the <:;A @niversal <ac et (<@<) remained an internetwor architecture as late as &'E' +4',. <@< was one potential means through which to improve on *A<, although as we see below, that role was later ta en on by #A<HI<. #his is one of many stories that call out for more research into the histories and the individuals involved. .et us now return to the story of the abovementioned wireless technologies to help e!plain the motivation that led to #A<HI< (as different from that which motivated <@<). #hese technologies led to wireless networ s that attached to the :;<:*=#, thereby e!posing the nature of the problems of supporting connectivity among heterogeneo us networ s. #he first step was ta en in 8ecember &'E1, when an IF< in Aalifornia used a satellite channel to connect to :loha*et through an :.O/: host in /awaii. #hus, the :;<:*=#, running the e!isting hostto-host *etwor Aontrol <rotocol, *A<, was now connected to a ground radio pac et networ , the :loha*et. #his was the first new networ to connect to the :;<:*=#. :loha*et had its own protocol and was wor ing independent of :;<:*=#, yet a gateway provided internetwor connectivity between the two. In &'E1 ;oberts e!tended the :;<:*=# to *orway over a leased line that :;<: had already installed to receive seismic data and then e!tended it to .ondon in the @nited

Ningdom. #his was the :;<:*=#9s first international connection. In .ondon <eter Nirstein then built a gateway to connect the :;<:*=# to a networ built with another protocol between the @.N. universities. #his was another case of different networ s $internetwor ing,% and as this function became an increasingly important focal point of :;<:*=# development, the networ came to be nown as the Internet to reflect this

growth. *A< was now handling the networ -tonetwor interconnect ion of :loha*et and the @.N. university networ , both of which were attached to the :;<:*=#. #he problems resulting from interconnect ed heterogeneo us networ s were becoming clear, and included the networ -tonetwor protocol conversion needed between any (and every) pair of networ s that were interconnect ed. It was clear that the combinatori al comple!ity of this pairwise protocol conversion would present considerable problems as the number of attached networ s scaled up. #A<HI< was soon to emerge as the response chosen to address these problems. :t 8:;<:&E in early &'E6, Nahn was the program manager responsible for, among other things, the ground

pac et radio networ and the satellite pac et radio networ . /e recogni3ed the differences between the :;<:*=# running *A<, and these two radio networ s. :s a result, he set out to design a scalable end-to-end protocol that would allow dissimilar networ s to communicate more easily. In the summer of &'E6 Nahn discussed his approach for dealing with this internetwor comple!ity with 7int Aerf of Gtanford who had considerable nowledge of *A<, since he had been a ey member of the @A.: software group involved in the *A< design. #ogether, they drafted a detailed design of a new protocol, the #ransmission Aontrol <rogram (#A<). #A< was to ta e over the *A<9s functions, but handle them in a more uniform mannerL it would allow applications to run over an internetwor while hiding the differences between networ

protocols by using a uniform internetwor protocol. #hey distributed this design at a computer communicati ons conference held at Gusse! @niversity in Geptember &'E6. (In October &'E6 ;oberts left I<#O to become A=O of #=.=*=#, the first commercial pac et switching networ carrier.) 2y &'E4, Aerf and Nahn fleshed out their design and published a definitive paper +5(, on #A<. @nderlying #A< was the ey idea of an open networ architecture that allowed pac et networ s of different types to interconnect with each other and for computers to e!change information end-to-end across these interconnecte d networ s. #his contribution by Aerf and Nahn was a critical step in the development
$%#$ was renamed :$%#$ in Larch 1('+ when the word 2:efense3 was
1'

prepended"

of the Internet. In &'E6U&'E4 8:;<: commissione d three independent implementations of #A<L Aerf at Gtanford @niversity, #omlinson at 22*, and Nirstein at @niversity Aollege .ondon. In addition, 8avid Alar of FI# wor ed on a compact version of #A< for the Tero! :lto personal wor station in the mid&'E(s and later for the I2F <A des top computerC 8avid ;eed, also of FI#, was wor ing on internetwor i ng among high-performance computers on .:*s for the .aboratory for Aomputer Gcience *etwor (whose wor was merged with the general #A< pro>ect in &'EB). In :ugust &'EB these implementati ons led to the first e!perimentati on using #A< to connect two different networ sL the pac et radio networ using Gtanford9s #A< implementati on, and the :;<:*=# using 22*9s #A< implementati

on. Following that, in &'EE, Nahn implemented the satellite reservation protocol ;oberts had designed, creating a second path from the :;<:*=# to the @nited Ningdom, sharing the capacity of a B4 bHs Intelsat I7 satellite broadcast channel among a number of ground stations in =urope and the =ast Aoast of the @nited Gtates. #his :tlantic <ac et Gatellite *et (later to be called G:#*=#) was the :;<:*=#9s second international connection. #his was the second twonetwor #A< demonstration. #hen a threenetwor demonstra$&

tion of #A< was conducted on *ovember 11, &'EE, when the pac et radio networ , G:#*=#, and :;<:*=# were interconnect ed to allow an Internet transmission to ta e place between a mobile pac et radio van at G;I and a @GAHIGI host computer (both in Aalifornia) via an intercontinental connection through @niversity Aollege .ondon. #his impressive feat was the first threenetwor #A<-based interconnect ion. #his first version of #A< only supported virtual circuits at the transport level (which

is fine for applications that re"uire reliable transmission) . 2ut it failed to support, among other things, realtime traffic such as pac et voice where many aspects of the session flow were more properly handled by the application as opposed to the networ . #hat is, realtime traffic called for support of an $unreliable% transport mechanism that would cope with missed pac ets, pac ets with errors, outoforder pac ets, delayed pac ets, and so on. #he use of unreliable transport support was already in use with *A<, prior to #A<C specifically, the early :;<:*=# IF< protocol allowed for

I=== Aommunications Faga3ine K :ugust 1(&(

H ISTORY OF C OMMUNICATIONS
unreliable transport by use of what was called type 6 pac ets (also nown as $raw% messages), which were introduced by Nahn in the 22* &)11 report. /owever, 22* was concerned that the uncontrolled use of these pac ets would degrade the networ performance, so they regulated the use of type 6 pac ets to be on a limited, scheduled basis. In &'E6U&'E4 8anny Aohen of @GAHIGI implemented a *etwor 7oice <rotocol (*7<) +5&, under :;<: support and re"uested 22* to allow him to use type 6 pac etsC with Nahn9s influence, 22* allowed this. Aohen9s real-time networ voice e!periments re"uired the ability to cope with unreliable data transport. #he early 7ersion & design of #A< in &'E4 did not support it, nor did 7ersion 1 when it was implemented around &'EE. It was around this time that pressure for supporting unreliable transport in #A< came from Aohen, now >oined by Dohn Ghoch and ;eed, and with involvement from Aroc er and 2ob 2raden. #hat is, they advocated modifying #A< such that type 6 pac et functionality would be supported alongside reliable data transport. Aohen convinced Don <ostel of this, and <ostel added a further concern, addressing layer violations, stating $We are screwing up in our design of internet protocols by violating the principle of layering. Gpecifically we are trying to use #A< to do two thingsL serve as a host level end-to-end protocol, and to serve as an Internet pac aging and routing protocol. #hese two things should be provided in a layered and modular way. I suggest that a new distinct internetwor protocol is needed, and that #A< be used strictly as a host level end to-end-protocol.% +51, <ostel then went on to describe how to brea #A< into $two componentsL the hop-byhop relaying of a message, and the endto-end control of the conversation.% : robust internetwor ing solution was no easy tas , and today9s #A<HI< was built with much e!perimentation on the ground laid by *A<. #hus, there was a clear call to cleave #A<, splitting the function of networ layer connectivity, which involved addressing and forwarding, from its transport-layer end-to-end connection establishment, which also involved flow control, "uality of service, retransmission, and more. #A< 7ersion 6 (&'E)) introduced the split into two components, but it was only in #A< 7ersion 4 (&')(, with an update in &')&) that we see a stable protocol running that separated out the Internet <rotocol (I<) from #A< (which now stood for #rans port Aontrol <rotocol) and was referred to as #A<HI<. #his version has come to be nown as I<v4. :long with the split into #A< and I<, the capability to support unreliable transport (i.e., type 6 pac et functionality) was included. #he formal name for this unreliable transport support was the @ser 8atagram <rotocol (@8<) +56,. In &')( the @.G. 8epartment of 8efense (8o8) declared +54, the #A<HI< suite to be the standard for 8o8. In Danuary &')6, #A<HI< became the official standard +55, for the :;<:*=#C after a short grace period of a few months, no networ was allowed to participate in the Internet if it did not comply with I<v4. Of course, Internet protocols never stop developing, and the &'') upgrade to 7ersion B dramatically e!tends the address space and introduces some significant security enhancements. It is still in the process of being deployed worldwide. Feanwhile, as the &'E(s rolled out, in addition to the :;<:*=# and #=.=*=#, other pac et networ s were being designed across the globe in this period. <eter Nirstein, in his earlier paper +5B, in this I!!! ,ommunications LagaKine /istory of Aommunications series, addresses much of the international wor , especially the @.N. story (to which we refer the reader for more details). :s a result of these national and international activities, an effort, spearheaded by ;oberts, was put forth that resulted in the International Aonsultative Aommittee on #elephone and #elegraph (AAI##) ;ecommendation T.15. #his agreed-upon protocol was based on virtual circuits ? which was to be the AAI##9s own e"uivalent of #A< ? and was adopted in &'EB +5E,. 8uring this period, the *etwor Feasurement Aenter (*FA) at @A.: was deeply involved in measuring, testing, stressing, and studying the :;<:*=#. 2ill *aylor and I published a summary of the tools used by the *FA as well as details of a wee long measurement and evaluation of the results in &'E4 +5),. In &'EB I published the first boo that described the :;<:*=# technology, including its analytical modeling, design, architecture, deployment, and detailed measurements. : summary of the :;<:*=# principles and lessons learned appeared in a &'E) paper +5', after almost a full decade of e!perience with the use, e!perimentation, and measurement of pac et networ sC this paper was part of a special issue on pac et communications which contains a number of ey papers

of that era +B(,. One

of the first measurements we made was to determine the throughput from @A.: to @AG2 in the initial four-node networ shown in Figure 1C note that there are two paths between these two nodes. Whereas only one path was tagged as active in the routing tables at any one time, we found that both paths were carrying traffic at the same time since "ueued traffic continued to feed one of the paths when the other path was tagged. :mong the more spectacular phenomena we uncovered were a series of loc ups, degradations, and traps in the early :;<:*=# technology, most of which were unintentional and produced unpredicted side effects. #hese measurements and e!periments were invaluable in identifying and correcting design issues for the early :;<:*=#, and in developing a philosophy about flow control that continues to inform us today. Foreover, it provided us, as researchers, a wealth of information for improving our theoretical models and analysis for more general networ s. In Duly &'E5 responsibility for the :;<:*=# was given to 8A:. #his terminated the systematic measurement, modeling, and stress testing that the @A.: *FA had performed for almost si! years, and was never again restored for the Internet.&)

It is outside the scope of this column to address Internet histories beyond those of its early period as the :;<:*=#. .i ewise, I have not done >ustice to the untold stories that abound, but I hope to have convinced the reader that many people contributed to its success. #his early history of the Internet, the first decade of design and deployment of the :;<:*=#, laid foundations on which today9s networ s depend and continue to develop.
1* @he wor of the AL, reOuired a strong

degree of cooperation from BBA since it was they who controlled any changes to the networ code and architecture" $t the AL,, each time we discovered a loc up, hardware problem, or other measured networ problem, we alerted BBA so that they would ta e corrective action" Over time we developed an efficient wor ing relationship with them, and errors were dealt with more expeditiously" It is worthwhile noting that the history of pac et networ s has met with institutional impediments to its progress, as have so many other technical advances over the course of history" In this case I have called out three with which I was personally involved5 $@7@Ps lac of interest in pac et switching, the researchers; reluctance to connect to the early networ , and the above1mentioned negotiation with BBA"

I=== Aommunications Faga3ine K :ugust 1(&(

$'

HISTORY OF COMMUNICATIONS
ACKNOWLE
DGMENT

&5, B.
et

I want to than 2radley Fidler, Fischa Gchwart3, and the reviewers of this column for their helpful comments and suggestions.

$e!ner al"< =T#e &a"t and F0t0re H!"tory of t#e Internet<>

Co,(0ter Net'ork"<>

IA&O%LS 1Oper1 ations %esearch<
Can.@Fe . E44E. &&&, $. Dle!nro%k< =Infor,at!on Flo' !n $ar)e Co,,0n!%a t!on Net"<> R$E G0arterly &ro)re"" Re(ort< MIT< A(r. ;26E. &&1, C. F. D0ro"e< M< S%#'artH< and Y. Ye,!n!< =Controll!n) W!ndo' &roto%ol" for T!,e: Con"tra!ne d Co,,0n!%a t!on !n a M0lt!(le A%%e"" En.!ron,en t<> #roc"

,ommun" $,L< .ol.
34< no. E< Fe . ;228< ((. ;4E@ 4?. &B, D. C. Red,ond and T. M. S,!t#<

REFERENC ES *A
&ARTIA$ $IST OF SE$ECTED REFERENCE S-

&rom Dhirlwind to LI@%!5 @he %7: Story of @he S$H! $ir :efense ,ompute r< T#e MIT
&re""< E444. &E, F. C. Cor atF< M. M. Da))ett< and R. C. Daley< =An E+(er!,e ntal T!,e: S#ar!n) Sy"te,<>

&&, W.

W. Ma""!e and C. R. Under#!ll< =T#e F0t0re of t#e W!rele"" Art<>

*th Symp" :ata ,ommun"<
O%t. 5@6< ;2?5< ((. 89@?3. &&6, A. R. Andre'"< =T#e De"!)n of a Me""a)e S'!t%#!n) Sy"te,I An A((l!%at!on and E.al0: at!on of Mod0la<>

;24?< ((. 68@8;. &1, H. A. Well"<

Direless @elegra1 phy and @elephony<

$&I#S ,onf" #roc"< .ol.
E;< ;26E SCCC< ((. 559@33. &), C. R. Ca%k"on< =Net'ork" of Wa!t!n) $!ne"<>

Dorld Brain<

Met#0en B Co. $td.< ;25?. &6, C. Ny%e and &. Da#n<

Operatio ns %esearch
< .ol. 9< no. 3< A0). ;298< ((. 9;?@E;. &', $. Dle!nro%k< =Infor,at!o n Flo' !n $ar)e Co,,0n!% at!on Net"<> &#.D. t#e"!" (ro(o"al< MIT< May ;26;. &&(, $. Dle!nro%k< =Creat!n) a Mat#e,at! %al T#eory of

&rom Lemex to Cypertext ? Nannevar Bush and the Lind;s Lachine<
A%ade,!% &re""< In%. ;22; &4, $. Dle!nro%k< =H!"tory of t#e Internet and It" Fle+! le F0t0re<>

I!!! @rans" Software !ng"< .ol.
SE:9< No. E< Mar. ;282. &&4, W. W. C#0< =De,0lt!(le +!n) Con"!derat! on" for Stat!"t!%al M0lt!(le+or "<> I!!!

@rans" ,ommun"<

I!!! Direless ,ommun"<
Fe . E44?< ((. ?@;?.

COM:E4I 645@642< C0ne ;28E. &&5, $. Dle!nro%k<

Gueueing Systems,

Nolume II5 ,omputer $pplication s< W!ley
Inter"%!en% e< Ne' York< ;286 &&B, $. Dle!nro%k<

al"<

=A D!)!tal Co,,0n!% a

t!on Net'ork For Co,(0ter" A!.!n) Ra(!d Re"(on"e At Re,ote Ter,!nal"<>

Lessage :elay in ,ommunic ation Aets with Storage<
&#.D. d!""ertat!on < MIT< Ca, r!d)e< MA< ;26E. &&E, $. Dle!nro%k<

$,L Symp" Op" Sys" #rinciples<
Aatl!n 0r)< TN< O%t. ;268. &&(, D. W. Da.!e"< =T#e &r!n%!(le" of a Data Co,,0n!%at! on Net'ork for Co,(0ter" and Re,ote &er!(#eral"<>

,ommunic ation Aets4 Stochastic Lessage &low and :elay<
M%Ara': H!ll< Ne' York< ;263. &&), &. Baran< =On D!"tr! 0ted Co,,0n!%a t!on Net'ork"> Rand &a(er &: E6E6< Se(t. ;26E. &&', D. Ho%#felder and &. Baran< $n

&&&,

#roc" I&I# Cardware <!dinburgh 1(/*> D;.

R. A. S%antle 0ry< &. T.< W!lk!n"on< and D. A. Bartlett< =T#e De"!)n of a Me""a)e S'!t%#!n) Centre for A D!)!tal Co,,0n!%a: t!on Net'ork<>

Oral Cisto1 ry< IEEE
H!"tory Center< E3 O%t. ;222< #tt(I77'''.! eee)#n.or)7 '!k!7!nde+.( #(7Oral: H!"toryI&a0 lJBaranKIn" t!t0teJForJt #eJF0t0re. ECJ;268.E B &1(, &. Baran et al"< =On D!"tr! 0ted Co,,0n!: %at!on"<> T#e RAND Cor(orat!on < Santa Mon!%a< CA< A0). ;263. &1&, D. W. Da.!e"< =&ro(o"al for a D!)!tal Co,: ,0n!%at!on Net'ork<> 0n(0 l!"#e d ,e,o< C0ne ;266< #tt(I77'''. ar%#!.e.or)7 deta!l"7 Nat!onal&# y"!%al$a or atory&ro(o "alForAD!) !: talCo,,0n! %at!onNet' ork &11, D. W. Da.!e" et

&&1,

#roc" I&I# Cardware <!dinburgh 1(/*> DE6.
R. A. S%antle 0ry< =A Model for t#e $o%al Area of a Data Co,,0n!%at! on Net'ork O /e%t!.e" and Hard'are Or)an!Hat!on <> #roc" 1st

$,L Symp", #roblems in the OptimiKatio n of :ata ,ommunica tions Systems<
&!ne Mo0n: ta!n< AA< O%t. ;5@;6< ;262< ((. ;?5@E43. &&6, C. C. R. $!%kl!der< =Man: Co,(0ter Sy, !o"!"<>

I%! @rans" Cuman &actors in !lectronics<
.ol. HFE:;< Mar. ;264< ((. 3@;;. &&4, C. $!%kl!der< and W. Clark< =On:$!ne Man: Co,(0ter Co,,0n!%at! on<> Spring

Joint ,omp" ,onf"<

Nat!onal &re""< &alo Alto< CA< May ;26E< .ol. E;< ((. ;;5@E?. &&5, T. Mar!ll and $a'ren%e Ro ert"< =To'ard a Coo(erat!.e Net'ork of T!,e:S#ared Co,(0ter"<>

&all $&I#S ,onf"< O%t
;266. $. Ro ert"< =M0lt!(le Co,(0ter Net'ork" and Inter%o,(0te r Co,,0n!%at! on<> $,L

&&B,

Symp" Operating System #rinciples<
Aatl!n 0r)< TN< O%t. ;268. &&E, See t#e fo0rt# entry 0nder =;268> !n #tt(I77'''.( a%ket.%%7 &&), $. Ro ert"< =Re"o0r%e S#ar!n) Co,(0ter Net'ork"<> C0ne 5< ;26?L #tt(I77!a5;4? 4?.0". ar%#!.e.or)77l oadJd/.0Ja( (let.(#(M f!leN47!te,"7 Re"o0r%eS# ar!n)Co,(0t erNet 'ork"7Re"o0 r%e: S#ar!n)Co, (0terNet'or k".d/.0 &&', =S(e%!f! %at!on" of Interfa%e Me""a)e &ro%e""or" for t#e AR&A Co,(0ter Net'ork<> De(art,ent of t#e Ar,y DBD7 /7O1 9:4323< C0ly E2< ;26?L #tt(I77'''.% ".0te+a".ed0 7 0"er"7%#r!"7D IAITA$JARC HIOE7AR&A NET7RFG: AR&A: IM&.(df &1(, $. A. Ro ert"< #tt(I77'''.H! (l!nk.net70"e r"7 lro ert"7SIA COMM22Jf!l e"7.5Jdo%0 ,ent.#t, &1&, =Interfa%

e Me""a)e &ro%e""orI S(e%!f!%at!o n" for t#e Inter%onne% t!on of a Ho"t and an IM&<> BBN Re(ort ;?EEL #tt(I77'''. !t: "a.er".or)7 (df7 n7!, (7BBN;?EE JCan;286.( df &11, S. Cro%ker< RFC ;< =Ho"t &roto%ol>L #tt(I77'''.f aP".or)7rf%" 7rf%;.#t,l &16, $. Dle!nro%k< =Model" for Co,(0ter Net'ork"<> Conferen%e Re%ord< I!!! I,,< Bo0lder< CO< C0ne ;262< ((. E;:2@E;: ;6. &14, T. T0)end< =UC$A to e f!r"t Stat!on !n Nat!on'!de Co,(0ter Net'ork<> UC$A &re"" Relea"e< 5 C0ly ;262. #tt(I77'''.l k.%".0%la. ed07$D7B! 7 RE&ORT7( re"".#t,l. &15, A. C. Cole< =Co,(0ter Net'ork Mea"0re: ,ent"L Te%#n!P0e" and E+(er!,ent "<> UC$A S%#ool of En)!neer!n) and A((l!ed S%!en%e< En)!neer!n) Re(ort UC$A: ENA:8;69 O%t. ;28;. &1B, S. Cro%ker< RFC 56< =&roto%ol Note">L #tt(I77'''.r f%ar%#!.e.or )7)etrf%.(# (Mrf%N56 &1E, S(e%! al "e""!on on t#e AR&ANET !n #roc"

,onf" #roc"<
.ol. 56< May< ;284< ((. 935@28I A. Ro ert" and B. D. We""ler< =Co,(0ter Net'ork De.elo(,en t to A%#!e.e Re"o0r%e S#ar!n)<> ((. 935@32. Heart et al"< =T#e Interfa%e Me""a)e &ro%e""or for t#e AR&A Co,(0ter Net'ork<> ((. 99;@68. Dle!nro%k< =Analyt!% and S!,0lat!on Met#od" !n Co,(0ter Net'ork De"!)n<> ((. 962@82. Frank< I. T. Fr!"%#< and W. S. C#o0< =To(olo)!%al Con"!derat!o n" !n t#e De"!)n of t#e AR&A Net'ork<> ((. 9?;@?8. Carr< S. Cro%ker< and O. Cerf. =Ho"t:Ho"t Co,,0n!%at! on &roto%ol !n t#e AR&A Net'ork<> ((. 9?2@2?.

$.

F.

$.

H.

S.

Spring Joint ,omp" ,onf"< $&I#S

&1),

S(e%!al "e""!on on t#e AR&ANET !n

#roc" Spring Joint ,omp" ,onf", $&I#S ,onf" #roc"<
.ol. 34< May< ;28EI S. M. Orn"te!n et al"< =T#e Ter,!nal IM& for t#e AR&A Co,(0ter Net'ork<> ((. E35@93. H. Frank< R. Da#n< and $. Dle!nro%k< =Co,(0ter Co,,0n!%at! on Net'ork Q De"!)n E+(er!en%e '!t# T#eory and &ra%t!%e<> ((. E99@84. S. D. Cro%ker et al"< =F0n%t!on: Or!ented &roto%ol" for t#e AR&A Co,(0ter Net'ork<> ((. E8;@82. R. H. T#o,a"< D. A. Hender"on< =M%ROSS Q A M0lt!: Co,(0ter &ro)ra,,!n ) Sy"te,<> ((. E?;@23. $. A. Ro ert"< =E+ten"!on" of &a%ket Co,,0n!: %at!on Te%#nolo)y to a Hand Held &er"onal Ter,!nal<> ((. E29@2?. &1', R. To,l!n"on< =T#e F!r"t Net'ork E,a!l<> #tt(I77o(en, a(. n.%o,7 Rto,l!n"o7ra y7f!r"te ,a!lfra,e.#t ,l &6(, N. A ra,"on< =T#e Alo#aNet Q S0rf!n) for W!rele"" Data<> I!!!

&61,

58< ((. ?9.

;284< E?;@

&6E,

$. A. Ro ert"< =A$OHA &a%ket Sy"te, '!t# and '!t#o0t Slot" and Ca(t0re<> AR&A Net: 'ork Info. Center< Stanford Re"ear%# In"t.< Menlo &ark< CA< ASS Note ? *NIC ;;E24-< C0ne ;28E. &66, $. Dle!nro%k and S. $a,< =&a%ket S'!t%#!n) !n a Slotted Satell!te C#annel<>

O. A. Cerf and R. E. Da#n< = A &roto%ol for &a%ket Net'ork Inter%onne%t! on<> I!!!

@rans" ,ommun" @ech"< .ol.
COM:EE< O 9< May ;283< ((. 6E8@3;. &6), D. Co#en< S. Ca"ner< and C. W. For)!e< =A Net'ork Oo!%e &roto%ol NO&:II<> ISI7RR:?;: 24< A(r. ;2?;. &6', C. &o"tel< =Co,,ent" on Internet &roto%ol and TC&<> IEN E< ;288. &4(, C. &o"tel< =RFC 86? Q U"er Data)ra, &roto%ol<> ;2?4L #tt(I77'''.fa P".or)7rf%"7rf %86?.#t,l &4&, A. &. D!nneen< =Ho"t:to: Ho"t Data Co,,0: n!%at!on" &roto%ol"<> A(r. 5< ;2?4L #tt(I77'''.( otaroo.net7!et f7!en7!en;9E.t +t &41, C. &o"tel< =NC&7TC& Tran"!t!on &lan<> RFC ?4;< No.. ;2?;L #tt(I77'''.rf %ed!tor. or)7rf%7rf%?4; .t+t &46, &. T. D!r"te!n< =T#e Early H!"tory of &a%ket S'!t%#!n) !n t#e UD<>

$&I#S ,onf" #roc"< .ol.
3E< NatSl. Co,(. Conf.< Ne' York< C0ne ;285< AFI&S &re""< ((. 845@;4. &64, S. $a, and $. Dle!nro%k< =&a%ket S'!t%#!n) !n a M0lt!a%%e"" Broad%a"t C#annelI Dyna,!% Control &ro%ed0re" <> I!!!

@rans" ,ommun"<
.ol. COM: E5<Se(t. ;289< ((. ?2;@243. &65, R. Met%alfe and D. Bo))" =Et#ernetI D!"tr! 0ted &a%ket: S'!t%#!n) for $o%al Co,(0ter Net'ork"<>

,ommun" $,L< .ol.
;2< no. 9< C0ly ;286< ((. 529@ 343. &6B, D. R. Bo))" et al"< =&0(I An Internet'or k Ar%#!te%t0re <> I!!!

I!!! ,ommun" Lag"< .ol.
38< no. E< Fe . E442< ((. ;?@E6 &44, T. Ry %Hn"k!< =&a%ket S'!t%#!n) *;289@ ;2?9-I A Canad!an &er"(e%t!.e<>

,ommun" Lag"< .ol.

38< no. ;E< De%. E442< ((. E;@E9. &6&, =T#e A$OHA Sy"te, Q Anot#er Alternat!.e for Co,(0te r Co,,0n!%at! on<> #roc"

@rans" ,ommun"<
COM:E?< no. 3< ((. 6;E@E3< A(r.< ;2?4< (. E6.

I!!! ,ommun" Lag"< .ol.

$&I#S &all Joint ,omp" ,onf"< .ol.

38< no. ;E< De%. E442< ((. E6@5;. &45, $. Dle!nro%k and W.

Naylor< =On Mea"0red Be#a.!or of t#e AR&A Net'ork<>

&4B,

$&I#S ,onf" #roc"<
.ol. 35< Nat!onal Co,(. Conf.< C#!%a: )o< I$< May ;283< AFI&S &re""< ((.

868@?4. $. Dle!nro%k< =&r!n%!(le" and $e""on" !n &a%ket Co,,0n!%a t!on"<>

&&(,

Co,,0n!%at! on"< .ol. 66< no. ;;< No.. ;28?< ((. ;5E4@E2.

#roc" I!!!<

#roc" I!!!<

S(e%!al I""0e on &a%ket

S(e%!al I""0e on &a%ket Co,,0: n!%at!on"< .ol. 66< no. ;;< No.. ;28?.

$6

I=== Aommunications Faga3ine K :ugust 1(&(