6 Introductum

Wide Web are prominent examples of informally created applications that became popular, not as the result of some central agency's mar­ Heat and ,-,UIH..4< keting plan, but through the spontaneous decisions of thousands of a.ndMeanings of Hacket(Switching independent users. In reconstructing the history of the Internet, I have been struck time and again by. the unexpected twists and turns its development has taken. Often a well-laid plan was abandoned after a short time and replaced by a new approach from an unexpected quarter.. Rapid advances, such as the introduction of personal computers and the invention of local-area networks, continually threatened to make existing network technologies obsolete. In addition, responsibility for operating the Internet changed hands several times over the course Of all the ARPANET's technical innovations, perhaps the most cele­ of its first thirty years or so. How, in the face of all this change and brated was packet switching. Packet switching was an experimental, uncertainty, did the system survive and even flourish? I believe that even controversial method for transmitting data across a network. Its the key to the Internet's success was a commitment to flexibility and proponents claimed that it would increase the efficiency, reliability, and diversity, both in technical design and in organizational culture. No speed of data communications, but it was also quite complex to imple­ one could predict the specific changes that would revolutionize the ment, and some communications experts argued that the technique computing and communications industries at the end of the twentieth would never work. Indeed, one reason the ARPANET became the century. A network architecture designed to accommodate a variety of focus of so much attenrion-wirhin the computer science community computing technologies, combined with an informal and inclusive was that it represented the first large-scale demonstration of the feasi­ management style, gave the Internet system the ability to adapt to an bility of packet switching. 1 The successful use. of packet switching in un predictable environment. the ARPANET and in other early networks paved the way for the The Internet's identity as a communication medium was not inher­ technique's widespread adoption, and at the end of the twentieth ent in the technology; it was constructed rhrough a series of social century packet switching continued to be the dominant .networking choices. The ingenuity of the system's builders and the practices of its practice. It had moved from the margins to the center, from experi­ users have proved just as crucial as computers and telephone circuits mental tp"normaJ" technology." in defining the structure and purpose of the Internet. That is what Many computer professionals have seen packet switching as having the title of this book, Inventing the Internet, is meant to evoke: not an obvious technical advantages over alternative nlt;:thods for transmitting isolated act of invention, but rather the idea that the meaning of the data, and they have tended to treat its widespread adoption as a Internet had to be.invented-and constantly reinvented-at the same natural result of these advantages. In fact, however, the success of time as the technology itself. I hope that this perspective will prove packet switching was not a sure thing, and 1'0\ many years there was useful to those of us, experts and users alike, who are even now no consensus on what its defining cbcaracteristics were, what advan­ engaged in reinventing the Internet. tages it offered, or how it should be implemented--in part because computer scientists evaluated it in ideological as well as technical terms. Before packet switching could achieve legitimacy in the eyes of data communications practitioners, its proponents had to prove that it would work by building demonstration networks. The wide disparity in the outcomes of these ePaul Baran at the Rand loomedrouover popular culture. .I'.he On the (Shute. Corporation in the United States and Donald Davies at the National R'Z-Safe (Burdick and Wheeler 1962)-both made mto Physical Laboratory in England. Baran was first to explore the idea, in tl at . 1 1960s-presented accounts of nuclear war around 1960; Davies came up with his own version of packet switching m()Vl1eSm ie eary '.. United States a few years later and subsequently learned of Baran's prior work. its aftermath. And in 1964, mOVIetheaters ' ., StlYLtl.,ge- presented a brilliant black comedy of Cold War paranOla, Dr. Davies was instrumental in passing on the knowledge of packet switch­ ing that he and Baran had developed to Lawrence Roberts, who was (Kubrick 1963). . 1 hili Dr St-rangeZavethough humorous, highlighted.the Vounera 1~ty. in charge of creating the ARPANET. This chain of invention and r, f , 1 d' ,, by a SOVIet 1 United States' communications'. channe s to lSIuptlon dissemination has become a standard element of origin stories about ::~ckwhich might make them unavailable just when th~ywere the Internet; indeed, it is easy to get the impression that packet switching simply took a detour through the United Kingdom before needed most. In the movie, a psychotic Air F.orce c..omman~elnk~me~ I · motion by 111VO mg a h t .. re-emerging, unchanged, in the United States to fulfill its destiny as J k D Ripper sets a nuclear .0 ocaus in •.. the underlying technology of the ARPANETl ac . of mutual assured destruction called "Plan R." ThIS plan- strategy .. ent president's authority to declare However, while Baran's and Davies's versions of packet switching which allows Ripper to crreumv . '1" had some basic technical similarities, their conceptions of what defined war-is specifically designed to compensate for a wa.rtime f~lUler~~ conllnan control. , and communications. In the mOVIe, an Air Fo packet switching and of what it was good for were very different. Much d,. general explains: of this difference was due to the strong political pressures that were brought to bear on computing research in tlwJJnited Kingdom and . Ian in which a lower-echelon commander may Plan R IS an emerl-?eflcy war p . k the normal chain of command in the United States. Large computer projects in both countries were k-if order nucle.ar retal;uon a~~r~~~:aw::t:~discourage the Russkies from any develQped in a context of government funding and control, and has been disrupte .. .. e .. W ihi t n as part ofa general sneak national leaders Saw computers as a strategictechnQIQgy for achieving h that they could knock out as mg 0 .. d I attackope and escape retaI'ia tion because of lack of proper command an contro. important political goals. But inthe very different policy contexts of

the United States and the United Kingdom, packet switching took on PI . R allows Ripper to launch a "retaliatory" attack even th.oug~1n~ different meanings for Baran, Davies, and Roberts. Packet switching firstan strike has actually OCCUlTed. In reality,(as t hems fil ' disclaimer.

was never adopted on the basis of purely technical criteria, but always ) th US Air Force never·ha.d any such strat..egy. Even before Dr. states, e . . d'ffl rent solu- because it fit into a broader socio-technical understanding of how data StrangeZave opened , the Air Force was explonng a very 1. e . te n networks could and should be used. . t the threat of a first strike: building.... a commumcauons sys 1 tron 0 . d h" er command that would be able to surviy,e a~attack an . so t at prop. 133 has Networking Dr. Strangelove: The Cold War ROQtsof Packet Switching an dcentro 1" could be maintained. As Edwards (1996, p... ) in the United States d C ld War defense analysts saw robust cornmurucauons documente, 0 f "Fl ible networks as a necessity in any nuclear con rontauon: exioie- As the 1960s opened, relations between the United States and the t .aregy required that political leaders contmue to commu- response s 1 . Th ,fore preserv- Union of Soviet Socialist Republics were distinctly chilly. The USSR nicate during an escalating nuclear exchange.'. . . . el e. ' . had launched its Sputnik satellite in 1957, setting off alarm in the mg,centra 1comman d and control-political... leadership,·'h buthi halso t United States over a "science gap" and prompting a surge of govern­ . ' .. communications., Iinks-s-achieved t e Ig es cedata and reconnaissan , . ., ... .•.,,' ,. was ment investment in science and technology. A series of events kept the l1lilitaryyrioTity." The need for survivable erally recognized the Co/4War 11 10 Chapter 1 White Heat

was a researcher at the Air Force's premier "think tank," the Rand Baran Corporation. for a new COllllllu,nilcatiOllS Founded by the Air Force in 1946 as an outgrowth of operations sUI'viv'abllltY with high capacity (ibid., pp. research efforts initiated during World War II, Rand (originally sioned a system would allow personnel to on VOICe RAND, derived from "research and development") was a nonprofit conversations or to use teletype, or low~speedcomputer corporation dedicated to research on military strategy and technology. terminalsunder wartime conditions. to this new was Rand was primarily funded by contracts from the Air Force, though a technique that Baran (1960, p. 3) "distributed communica~ it served other government agencies as well. It attracted talented tions." In a conventional communications system, such as the tele­ minds though a combination of high salaries, relative autonomy for phone network, switching is concentrated' and hierar~hic~l.Calls g? researchers, and the chance to contribute to policy decisions of the first to a local office, then to a regional or national switching If highest importance (Baran 1990, pp. 10, 11). Edwards (1996, p. 116) a connection beyond the'Iocal area is needed. Each user is connected notes that "Rand was the center of civilian intellectual involvement in to only one local()ffice, and each local office serves a large number defense problems of the 1950s, especially the overarching issue of users, Thus,destroying a single local office would cut off users nuclear politics and strategy." Rand's role was visible enough to be from the network. A distributed system would have many switching reflected in popular culture-for example, the fictional Dr. Strange­ nodes, and many links attached to node. The redundancy would love turns to "the Bland Corporation" when he needs advice on make it harder to cut off service to users. nuclear strategy;' Because its approach to systems analysis emphasized In Baran's proposed system, each of several hundred switching quantitative models and simulation, Rand was also active in computer nodes would be connected to ' nodes by as many as eight lines science research (Edwards 1996, pp. 122-124). (figure 1.1). Several hundred multiplexing stations woul~pro~idean In 1959 a young engineer named Paul Baran Rand's com- interface betweea the users and the network. Each multiplexing sta­

puter science department, Immersed in' a rornor,,,tp E·"I'r,~afocused tion would be connected to two or three switching nodes and as many as 1024 users with data terminals or digital telephones. on the Colct Wal~Baran Soon fl.ev~l()ped~fl in survivable switching was distributed among all the nodes in network, so communications, which he feltwoulct cte~reasetel;nptation of rnili.. tary leaders to launch a preemptive first strike: knocking out a few important centers would not disable the whole network. To make the system even more secure, Baran (I964a, volume Both the US and USSR were building hair-trigger nuclear ballistic missile VIII, section V) planned to locate nodes far from population systems If the strategic weapons command and control systems could be .... centers (which were considered military and he designed the more survivable, then the country's retaliatory capability could better allow ~t to withstand an attack and still function; a more stable position. But this was multiplexing stations with a wide margin of excess the not a wholly feasible concept, because long-distance communications networks assumption that attacks would cause some equipment to fail). Baran at that time were extremely vulnerable and not able to survive attack. That added such military features as cryptography and a was the issue. Here a most dangerous situation was created by the lack of a that would allow high~levelusers to preempt messages from lower- survivable communication system. (Baran 1990, p. 11)5 level users. Baran was able to explore this idea without an explicit contract from To move data throllgh the network, Baran adapted a technique the Air Force (ibid., pp. 12, 16), since Rand had a considerable amount known as "message or "store-and-forward switching." A of open-ended funding that researchers could use to pursue projects common example of message switching is In a they deemed relevant to the United States' defense concerns." message switching system, each message (e.g., a letter) is labeled with Baran began in 1959 with a plan for a minimal communications its origin and its destination and is then passed from node to node system that could transmit a simple "Go/No go" message from the through the network. A message is temporarily stored at each node president to commanders by means of AM radio. When' Baran pre~ (e.g., a post office) until it can be forwarded to next node or the sented this idea to military officers, they immediately insisted that they final destination. Each node uses the "rlrlr,,,',,,information 12 Chapter 1 White Heat and Cold War 13

I; I I. to determine the next of mute. In message SW.LLUl· I I· ing came into use in telegraphy: a message was stored on paper tape MI' ~~~~~---J at each intermediate station before being transmitted to the next ----- s -______~/ ":.7 s I: station. At first, telegraph messages were switched manually by the I: telegraph operatorsc.however, in the 1960s telegraph offices began to I: I. use computers to store and route the messages (Campbell-Kelly 1988, I: p.224). For the postal and telegraph systems, message switching was more efficient than transmitting messages or letters directly from a source to a destination. Letters are stored temporarily at a post office so that a large number can be gathered for each delivery mute. In telegraphy,

message switching also addressed the .•uneven. flow of traffic on the expensive long-distance lines. In periods oflight traffic, excess capacity :00'­.. I: was wasted; when the lines were overloaded, there was a.riskthat some I; I' messages would be lost. Storing. messages at intermediate stations I: made it possible to even out the flow: if a line was. busy, messages could I: be stored at the switch until the line was free. In this way, message

••••••••------­• .... •• •• 0 ...... switching increased the efflciency, and hence the economy, of long­ distance telegraphy} Besides appreciating the efficiency offered by message switching, Baran saw it as a way to make his system· more survivable..Since the nodes in a message switching system act independently in processing the messages. and there are no preset routes between nodes, the nodes I; I' can adapt to changing conditions by picking the route that is best at I: any moment. Baran (1964b, p. 8) described it this way: "There is no I' I: central control; only a simple local Touting.policy is performed at each - _ _ _ _ s t:":::-::------.J S:.. 7.-:-.:-: .7.:-..-: .:-.:- .:-. ~_ node, yet the over-all system adapts." This increases the ability of the system to. survive an attack, since the nodes can reroute messages I: I

~user or "subscribe," around non-functioning parts of the network. Baran realized .that , .. - ....~emote subscriber'$. ~ survivability depended on more than just having redundant links; the ~ ~ Non-secure terminq! cree -s;> Multipluing station nodes must be able to make use .of those extra links. "Survivability,"

LoCOIS\lOscribers o SWitching ned. Baran wrote (1964a, volume V, sti,ction I), "is a function of switching Sock-up link to switching nOde Infer- switetting node flexibility." Therefore, his network design was characterized by distrib­ horizontal cnd M . Primory route from vertical rou1" uted routing as well as distributed links. . multlple.lng station tal from SWitching node ~ •..••. •. Diogonel routes second bock-up Departures from Other Contemporary Systems Paul Baran was not the first to propose either message switching Figure 1.1 or survivable communications to the military. Systems of both types Paul Baran's design featuring highly connected switching nodes. Source: Baran 1964a, volume VIlI. already existed or were in development. A look at the state of the art in these areas makes it easier to see what aspects of Baran's ideas were of of

15 the

the the

this the

that

sys- with

that

each

from

ways. Lines

"The

traffic

coor-

major

major

in

survive

AUTO­ a

existing

was

that

make

Network

virtues

rerouted,

switcbing switching

operatio~s

with

nodes. " with lines.

Lines

the

from

be

method

a

the connection ten switching

"plays

of

to

other

whereas

departure throughout would

Voice

throughout significant

single

network.

stressed

best

away distributed

a

with had

charzed

conneenons

269).

thousand

toll

in'

magazine

.... on

lines,

sharp

placing reliable

the

lights, center

message

all with its

links

a

in two

architects,

a

services 266 227)

traffic

publicity

built

a

and

distributed at

If

is

not

pp.

Along

control AT&T's

ordinary

part

having

hundred

warning

into

was

system's underground,

rapid

represents

in 1968, p.

served the

the

nodes

1982,

network

AT&T's

represented few

at

the

in

the

went'

sUl'viv:abilityfeature,however,

was

from~T&T's

to often had

operations.

It

voice

requirement sought

for

tht:?operators·

(Gorgas

COll11UIIWaieaw:ms

typically every used

AUTOVON.

polygrid

in system monitored

main

(rather

top nodes

sites, of

switches).

VON

was

routes (Schindler

operators

for

system

offer

differed

The node

military

The according contact

a

network.

those

1964 to AUTOVON

structure now

AUTO node

arranged

of

central

economy,

change

operators

of

then

the

targets;

special the

manually:

was

can decentralized

system

noted:" one controlled

survivability "hardened"

a approach

and

Survivability

and

a few network,"

3)

April

communications."

toll

in control where

telephone

the

h connected

.and had done

although

in

in

grid

hierarchical

neltw()rk

Baran's

De:scx:ibing

it was de'ClS],On First, system, center, technology defense [evels, poly flexibility redundant, re!lrul'lr through(.mt mroug

l"llt~tr(Dp()lit:an

(:ellteJ['S

.~Yll1an

of

to

or

be

of

and

con­

with

net­ and pos­

par­

to be

p. 6)

have mes­

These

early trans­

bottle­

would

radical

in con­

In

systems to

connec­

propos­

Depart-

messages was message essential

built switching

from large tended

should

little time

inefficient.

the store-and­

the

existing

it I). of

distributed route

Department

distributed

in examples

be

civilian

used

the

1960s.

waiting

a (1964b,

only be seem

in

Department

of have

and

and

to

every

that

being

of cite

for

getting

depart

"distributed spend

multiple

and

large

connection

the

would variety section those

several were early

were

idea 24) them up messages had to

of

a

by

with could

(or a Baran

..

could would differed

V;

systems.

fastend-to-end of

eliminating

one

of

the

switches was a

the

that

than

would

As

member

new to

system

operated

such

with ended

linked was

benefit most

system

development cumbersome switches

system publications

(1990, p

of volume described

systems

as a network

therefore, examining speeds,

intelligence,

and

proposed

shared

backlogs a

his

the

techniques,

cheaper

there

systems.

nodes

messages

Baran's

nothing were opportunities

messages

systems in

network

actual

user,

with obvious

inexpensive they other

Baran

1964a,

that that

served

and

the

computers particular,

data

areas.

network."

already

low-storage were

the argument data

artificial

the

he

indeed, designed

high-speed

In

these

to systems

whereas

small.. broadcast

practice the under

hold

transmission However,

Therefore,

had

systems (Baran

understanding

charged Though

the

switching

capacity simpler

in to

when certain

why he saw

decentralized

Baran; systems Baran's

convinced

and

Besides switching

in

nodes;

IV). of

low simple

this

existing higher

proposal.

and "distriburedcommunications''

high-speed,

through

however,

was much

the

systems. network

existing of

switches.

the planners

AUTOVON,

mathematics,

of

design, message

destination,

the capacity

although

1 such nodes. distributed

a

at

committee

the

the

use predated and

much practice

switching

section were store-and-forward

real-time. messages

was

store-and-forward

network

Baran Baran's

but faster, a to

the at

the

from

a out, in its had

of every

the

communications

also

working

build

discovered

military

innovative of

that

concept

at Chapter

up

storage stored

This

on using to be to

(ibid.,

individually

from

idea

Defense

not

Message

The

Most

14 really temporary

Defense, Baran of proposed

systems build large forwarded, necks complex. built delivered nodes temporary

pointed forward being seem

sible to mission challenge works") the

military ticular, tions based

ways communications, entailed going sages 1960s.

als, communications 16 Ch(lpter 1 White Heat Cold..War 17

Nodes would be individually responsible for determining routes, and Trivision. which ran the service would do so automatically without human intervention: "The.intelli­ system, tended only with analog gence required to switch signals to surviving links is at the link nodes te_:tn:\;otogy,and they doubted Baran's claims that an all-digital system and not at one or a few centralized switching centers." (Baran 1960, tr.uls,cend the' well-known limits on the number oflinks per call p. 3) Clearly such a. system would be more survivable than one p.18).lQ Whereas inAUTOVON there was a max~mu~l

dependent on a single operations center-which, Baran noted, "forms .~~l",,;¥\.ulinks in any route, Baran's simulations of network routmg m a single, very attractive target in the thermonuclear era" (1964a, vol­ version ·of as .23 links between ume V, section II). \~n(tpolmts(Gorgas1968,223; BaranI964b,p.l, 11). Ev~de~tly, One implication of Baran's design was that the nodes would have to position outside .the community of analog commumcanons have enough" intelligence" to perform. their own routing-they would and his awareness of the potential of computer tech­ have to be computers, not just telephone switches. This brings us to made it easier for him to question the accepted limits. lIe had Baran's second departure from the AT&T approach: Baran envisioned in analog telephony, and his training and background in

an all-digital network, with computerized switches and digital trans­ computing made it easier for him to envision an ~ll-d.igitalsystem as mission across the links. The complexity of routing messages would of achieving his goal of distributed commumcatron. require computers at the nodes, since the switches would have to AJl;d Baran's system departed from traditional telephone company be able to determine, on their own, the best path to any destination, Pl';actlce in other ways that show the effect of Cold Warmilitar~con­ and to update that information as network conditions changed. Such his design assumptions. For instance, AT&T tried to computerized switches had never been designed before. "These prob­ increase the reliability of the phone making each lems," Baran acknowledged (1964b, p. 6), "place difficult require­ reliable as possible, fee would

ments on the switching. However; the development ofdigitalcomputer providelines that werespeciallyBell Labs these messages were all sent in digital form-as a series o~binary had only begun developing its T'Ldigital trunk lines in 1955, and they numbers ("hits")~meantthat the information could be manipulated would not be ready tor commercial service in the Bell System until in new ways. Baran proposed that, rather than sending messages of 1962 (O'Neill 1985).9 varying •sizes across the network, messages should be' divi~ed'i~to Baran's system would push contemporary switching and transmis­ fixed-size units that he called "message blecks.": The multiplexing sion technology to their limits, so it is understandable that contempo­ stations that connected users to the for rary experts reacted skeptically to his claims. The engineers in AT&T's dividing oUl:goi,ng messages 18 Chapter 1 White Heat and 19

message could be sent as a single block; longer messages would rprUll1rp multiple message blocks. The multiplexer would add to each block a header specifying the addresses of the sending and receiviI}g parties his as well as other control information. The switching nodes would use fundamental goals of the system. the header information to. determine what route each block should benefits. Baran was determined t~k~to its destination; since each block was routed independently, the than dIfferent blocks that made lclp a single message might be sent on message s\vitchirlg systems, and he

different routes. When the blocks reached their destination, the local ~Si.·aVvaI;ethat the switching would have to be simple in multiplexer would strip the header information from each block and ~ij,t~~ertoue U()UI fast and inexpensive. The use'of fixed-size packets reassemble the blocks to form the complete message. This idea would variable-size messages could simplify the design of the eventually be widely adopted for use in computer networks; the mes­ Another advantage for the military was that breaking sage blocks would COme to be called "packets" and the technique packets and sending them different routes to their "packet switching."ll make it harder for to eavesdrop on a conver- For all its eventual significance, the decision to transmit data as the biggest potential reward was and flexible packets was not the original focus of Baran's work. As the title of his data. "Most importantly,"wrote Baran (I964b, p. 6), eleven-volume work On Distributed Communications indicates, Baran blocks permit many simultaneous users, each with began with the idea of building a distributed nerwork-s-an idea that "!!iJUueIV dHielrerlt bandwidth' requirements[,] to economically share a h~~already been identified with survivability by people working. in yr~~t?oa(l··palnanetwork made up of varied data rate links." In other military communications (Baran I964a, volume V). In describing the ••..·.·.·.~v!or<1s,Ud.CKC:Jt switching allowed a multiplexing system, Baran tended to stress the idea oflink redundancy, rather than ts.JlarJlng of a single communication channel by users). othe.r elements such as packet switching.12But as. he developed the the conventional telecommunications systems of the early 1960s, details of the system, the use of message blocks emerged as a funda­ usual form of multiplexing was by frequency division: each caller mental element. By the.time he wrote the final volume of the series, be assigned a particular frequency band fOTtheir exclusive use Baran had changed the name he used to refer to. the system to reflect tile duration of their connection. If the caller did not talk or send the new. emphasis: '.'While preparing the draft of this concluding continuously, the idle time would be wasted. In an alternative num~el~It bec.ame evident that a distinct and specific system was being I,lle:th()d, called "time division multiplexing," time is divided into short described, which we have now chosen to call the 'Distributed Adaptive jutE'"'",al,, and each user in turn is given a chance to transmit data for Mess~geBloc~~etwork,'in order to distinguish it from the growiIig duration of one interval. Only users' with data to transmit are set of other distributed networks and systems." (Baran 1964a, volume oJfered time slots, so no slots go idle as long as anyone has data to XI, sec~iOl:I) What, then, was so important about packet switching? transmit; this makes time division multiplexing more efficient for What did It mean to Baran and his sponsors? usage situations where bursts of information alternate with idle peri­ .Transmitting packets rather than complete messages imposed cer­ ods.Since/ computer data tends to, have this "bursty" characteristic, tam costs on the system. The interface computers had to.perform the Baran (1964b, P: 6) felt that time division was a more "natural" form work of dividing users' outgoing messages into packets and of reas­ multiplexing for data transmission. And since the time slot accom­ sembling incoming packets into messages. There was also the overhead modateda fixed amount of data, Baran believed that the use of of having to include address and control information with each packet fixed-size message blocks was a prerequisite for time division multi­ (rather than once per message), which increased the amount of data plexing. Thus, he associated packet switching with time division multi­ that had to be transmitted over the network. And since packets from plexing and its promise of efficient data transmissien.U a ~inglem:ssage cou.ld take different routes to their destination, they Packet switching would also make it easier to combine links having might arrive out of sequence, which meant that there had to different data rates in the network. The data rate is number of bits 20 Chapter 1 White Heat and Cold War 21

per second that can be transmitted on a given link. In the conventional that Baran had described, reluctaatlv telephone system, each caller is connected at a fixed data rate, and the prepesal rather it executed data must flow into and out of a switch at predetermined rates. Wi~h sums of money and perhaps discredit Baran's ideas packet switching, data flowing intoa switch can be divided a1nong the pp. 33-35),!5 outgoing links in a variety of ways, rather than having to be sent out e proposed network was never built.ilsaran's.ideas were at a fixed rate. This would IlJake it easier for devices.transmitting data disseminated among researchers interested.in new communi­ at different rates (computers and digital telephones, tor instance) to technologies. Following; Rand's standard practice, Baran pre­ share a link to the network. The system could also take advantage of his work to various outside experts for comment as he was new media, such as low-cost microwave transmission, that had differ­ ping his ideas.I" Eleven volumes of reports published in 1964 ent data rates than the standard phone company circuits. Th(:llJgh widely distributed to individuals, government agencies, Rand packet switching made the system more complex in some respects, in 'orylibI'aries, and other people working in the field. The first other ways it made the system simpler and less costly to build. was also published as an article in the March 1964 issue of In sum, packet switching appealed to Baran because it seemed to Trl't?t!,ac,ci017Son CommunicationsSystelns, and an abstractappeared meet the requirements of a survivable military system. Cheaper nodes August 1964 issue of IEEE Spectrum (a magazine for electrical and links made it economically feasible to build a highly redundant cowputing engineers with an estimated circulation of 160,000),17 (and therefore robust) network. Efficient.transmission made it possible also lectured on his work at various universities (Baran 1990, for commanders to have the higher communications capacity they It is not dear how many researchers wereammediately

wanted. Dividing messages into packets increased security by making [l.t;;ll1..C;;U by Baran's ideas through thesechannels. Most academic it harder to intercept intelligible messages. Packet switching, as Baran erscientists were not concerned with the survivability of come understood it, made perfect sense in the Cold War context of his . ns, and they may not have seen the applicability of Baran's proposed system. ch to their own interests. Several year-slateI', however, his work begin to receive wide attention as the technical founda- The Impact of Baran's Work ARPANET. Curiously enough, the connection between For a brief time after its publication in 1\;)61, it seemed that Baran's American networking efforts would be made via a laboratory On Distributed Communications might Soon become the .blueprint for a nationwide distributed packet swi~chingnetwork, In. Avgust of 1995, Rand officially recommended that the Air Force proceed with research #i()r./fingPacket Switching in the White Heat: Networks and

and development on a "distributed adaptive message-block network." /!Il(~ti,(~n(i~lismin the United Kingdom Enthusiastic about the proposal, Air Force representatives sent it for review to the Defense Communications Agency, which oversaw the the early 1960s, while.the United States was caught up in the Cold provision of military communications services (Baran 1990, Attach­ United Kingdom was experiencing political upheaval of a ment 2). The DCA was one of many agencies that had been created type. Just as the America9s were worried about a "science in an attempt to bring military operations underthe central control of between their country and the USSR, so there were widespread the Defense than aUowingeach the Department of rather of armed .·.~ars,·ln,tne United Kingdom ofa"technology gap" with the United services to build its own systems.t! In accordance with this centralizing ""fatE'S. Harold Wilson was elected leader of the British Labour Party strategy, the DoD administration made it dear during the review 1963, at a time when that party,.and much of the general popula­ process that any new network would be built not by Air Force contraj.;­ Ujon.;·te,ltthat-the UK was facing an economic crisis. Politicians on all tors but by the DCA, which had no expertise in digital technology. that the UK was falling behind the other industrial Baran and his Air Force sponsors, doubting that the DCA would be exploitation of new technologies, that there was a "brain 22 Chapter 1 White Heat and Cold War 23 drain" of British scientists to other countries, and that the country's that for technological backwardness was at least partly responsible eCO­ nomic malaise (Coopey and Clarke 1995; Edgerton 1996, pp. Research Development Corporation, which gave develop­ ds to corporations that wanted to commercialize government Wilson addressed the technology issue head.on.in a. speech.to..the Labour Party's annual conference at Scarborough pn 2 October 1963. h, and by using governmenLcontracts to encourage the intro­ Calling on labor and management to join in revitalizing British indlls­ ion of new computer products (ibid., p. 63). In addition, Mintech try, Wilson stressed the importance of keeping up with the ongoing Industrial Reorganization Corporation were responsible for scientific and technological revolution, and hei1'lvoked a stirring vision ing British corporate mergers to large companies, such as of a new United Kingdom "forged in the white heat of this revolution" ational Computers Limited, which would supposedly have the (quoted in Edgerton 1996, p.56). The speech.created a sensationin of resources to compete internationally (Hendry 1990, the British media, and Wilson was praised in. newspap~rsaCross the Wilson 1971, p. 63). In 1965 Mintech also took over a political spectrum for capturing the concerns ()f the times and remak­ ernment initiative called the Advanced Computer Techniques Proj­ nt- ing Labour's supposedly anti-progress image.J~When Labour came to , which had been set up in 1960 to help spin off governme power in the 1964 general election, Wilson was eager to act on. his ~p!on:s()r'edcomputing research to industry. Under Wilson, computing vision by implementing a new economic and technological regime for Jieseal:ch was expected to serve economic aims, and the possibility of the United Kingdom. 1I0Vel:nrnellt intervention was always present. Wilson's plans included reversing the "brain drain" by training more 'j;;.!~rlCofthe British scientists who tookthe lead in computing research

;'~~ls,Rona]ld scientists and giving them the status and the facilities that .would •. W. Davies of the National Physical Laboratory in Ted­ persuade them to stay in the United Kingdom, by rationalizing exist­ a suburb of London.. The. NPL-established in 1899 to ing industries and creating new high-tech indllstries, and by shifting physical constants; to standardize instruments for resources from unproductive defense and "prestige" areas (such as measurements, and to .perform similar activities involving aerospace and nuclear energy) to cornmercialapplications.Jo oversee modem and a telephone line, and access computer can serve many terminals at once, it will spend less time idle the~ervice'scomputers for an hourly rate. Many people thought that and more time doing productive work, which increases the efficiencs-« time sharing represented the future of interactive computing, since and therefore the economically feasibility---of interactive computing. 26 Chapter 1 White Heat and Cold War 27 few if any anticipated the advent of small, inexpensive "personal" U1 computers in the late 1970s. Davies became interested in time sharing during a 1965 trip to the United States. He had gone there to participate in a computing con­ ference being held at MIT, and he took the opportunity to visit several t,switching in Davies's System American computing research sites (Davies 1986, pp. 4-5). Both the Paul Baran, Donald saw packet switchingwould allow conference and his site visits made it clear to him that interest in and to share a communication But Davies knowledge about time sharing were much more widespread in the (Y"#l~ll'te(ltJllatefficiency for a in his United States than in the United Kingdom. When Davies returned to the comlnunications it the National Physical Laboratory, he decided to organize a seminar on rn-ovide afford- time sharing to disseminate these ideas to the British computing com­ munity. The seminar was held in November of 1965, and a number >.;:>,u~.•,....., .." of1966,his nel:wor~ of British and American researchers were invited. time, to an enthusiastic audience of people It was during these discussions that Davies became aware of a widely //~.~g,:telec.orlunun.lcaIUc'ns:and the military. perceived obstacle to interactive computing: inadequate data commu­ /1I!iritisb:MinJlstr'y of Defence gave Daviesthe surprising packet nications. In early time sharing systems, the terminals had been already been invented a few years earlier by an Ameri- directly connected to the computer and were located in an adjacent fE>aran}. The fact that the military about this earlier terminal room. As time went on, people began locating terminals at when Davies did not undersceres the very different some distance from the computer itself, either for the user's own in which packet evolved in the two countries. convenience or, in the case of commercial time sharing services, to offer foremost concern had was under- access to customers over a wide geographic area. Distant terminals his use of terms like " "attack could be connected to the comPllter using qiahllP~otelephone links of kill" in hostile conditions under and modems, but 10ng-diSctance tel~phpneiS()I),nectionswere very his system was expected to operate (Baran 1964b, p. 2). Davies, expensive, and for data transmission they also inefficient. Com­ LH!;:;U'LU'CI hand, did not view packet switching as a way to make the puter messages, as noted earlier; tend to come in bursts with long survivablet after reading Baran, he commented that "the pauses in between, so compllte:r users paid dearly for telephone con­ connected networks there considered" "not needed in a nections that were idle much of the time. The high cost of communi­ (Davies 1966b, p.21). 23 Davies thought the pressing cations put pressure on users to Work qllickly, sacrificing the user for a network that could serve the users of commercial time friendliness for which time sharing had been invented." Davies had a services. This assumption is evident in his plan to survey long-standing interest in switching techniques. Ashe thought about businesses' data communications requirements (Davies 1968a). It also the data communications problem, he came up with the idea that a in Davies's efforts to make the system easy use. In his new approach to switching might offer a solution (Davies 198(:i, pp. 6­ national.lletWork, "Afurtller aim requirement 7). He knew that message switching was used in the telegraph system keep in mind constantly the use of the system to make efficient use of lines, and he believed that by adapting this simple for simple jobs. Even though there is a communication system technique to c()mputer communications he could achieve similar ~l1v I>lC:IU. Rather, in its entirety before the next message could begin. In a packet switch" compete ing system, time division multiplexing would allow users to take turns business transmitting portions of their messages. If a user had 4 short message, compete with such as a single command for a time sharing system, the whole message mnovatrve computer could be sent in the first packet, while longer messages would take several time slots to transmit. This way, the user with the short message of a national packet would not have to wait behind users with long messages (Campbell­ Pl'O\'ld,e;;lnexpenslve data communica­ Kelly 1988, p. 226). This kind of fairness was appropriate for asys­ envisioned the net­ tern where computers were serving the everyday needs of civilians, J::»llSlJneilSand recreational rather than transmitting life-or-death messages through a command transactions, hierarchy. even online betting Ultimately, Davies thought, packet switching technology could be" hl!,?;J4,·capal:ltytelephone come a commercial product that would contribute directly to Harold K,lllg(lprn; the proposed Wilson's plan to revitalize the British economy. In a 1965 proposal to not have the General Post Office build a prototype for a national packet as re(lu.r1Qllll1k 1with switching network, Davies (1965, p. 8) wrote: J?hysical Lal)or;ltQIY",.•,,~.... (leSII~nettal1et:wol'k a !dvnarnic. tributed rOlltinlg" Such an experiment at an early stage is I}ee~tdt9deyelop the knowled~eof pendently to current conditions in these systems in the GPO and tile British computer ~~.t:icommunicarions industry. ... It is very important not to find ourselves forced to buycomput­ would be connected by high-sp~edtelephone lines so provide fast ers and softwarefor these systems fl"om [the] USA.We could, bystartir~gearly r~sponsefor int~ractivecomputing:Us~rswould attach their comput- enough, develop export marketsJ? terminals, and printers to the nodes through dedicated interface c::omputers at local sites. Davies (1968(.:, p. 7) reiterated the need to compete with the United Davies was convinced that a data communications infrastructure of States in 1968, when he compared a proposed Mintech network with the sort he was proposing would be necessary to keep the United the planned ARPANET: Kingdom competitive in the information age, and he did not doubt

The proposal resembles the ARPAnetwork being planned.... The sponsors I.h4t such. a n~tworkwould som~daybllilt. l:Imvevel~the NPL did of that project believe it will "spearhead" a new kind of data communkation not have the resources or the authority to build such a large network system to be developed on a nation-wide scale. on its own. This authority belonged to the General Post Office, which A Mintech network would go beyond the present ARPAplans by providing ran the national postal and telephone networks, but managers there for a variety of terminals as wellas computer to computer communication. To be useful as a "spearhead" project it would need to be started soon and had little knowledge of or interesr in data communications. Since planned with as short a time scale as possible, coming into operation well Davies felt there was no hope of convincing the GPO to collaborate on before a national network. a national network, he decided that a small in-house experiment would be the only feasible alternative. In the summer of 1966 he made For Davies, the network was not only a communications t0o.l;it was a second, much more modest proposal to build a prototype network also a way for British researchers to apply the "white heat" of ~ci~ntific at the NPL. This network, named "Mark I," would serve as a demon­ innovation to counteracting American domin~ncein th~computer market. . stratien of packet switching, advance the state of knowledge in me field, and support the operational of the NPI~s Davies's concern with economics and user friendIinesslll1dTrscores scientific and administrative personnel. the national context in which he conceived the idea of a packet switch- Ht'hiteHeat Gold War 31

a dlevelopnlerlt team headed

"Y""'~''''''''I:l,arl)er. ~.ogert)lcan,tlel)UI'y .."..", Jeu,- technical leader, Keith oversaw hardware development, and Peter Wilkinson was in f software development (Campbell-Kelly 1988, pp. 228-229). they brought skilland enthusiasm to the project, the members PL team had to struggle agail'lst technical aIt:9 fimmcial con- . The Mark I was originally designed to have three packet g nodes, but funding constraints reduced the nUlnber to one. ly on~node, the NPL team would not get the opportunity to certain issues-such as congestion and routing-that a multi­ network would raise. Davies hoped that simulation studies could co~pensatefor the lack of direct experience with a full-scale network , p. 229). The team had chosen for the Oc0dea computer, made re English manufacturer Plessey, had, heen designed spe­ lly for data communicatio er the NPLgroup bad spent a designing the Mark I arou this ll;lachine, it abruptly with- n from the market, and the team had to new computer. chosen computer was th~Honeyw~ll Hone)"vell 516 was installed at the 1969, and over the next two years user services were added to the network. The Mark I had about sixty lines that provided access to a DEC PDP-8 computer and two mainframes. Through the network, NPL researchers could have remote access to computers for writing and running programs, for querying a database, for sharing files, for special services such as a "desk calculator," and for "communication between people" (Davies 1966a, pp. 1-2). The system also included a file server and a "Scrap­ Figure 1.3 Q?ok" application that provided document editing and communica­ Donald Davies's proposed network for the United Kingdom (from archival tion tools (Campbell-Kelly 1988, p. 236). copy). ,..Most of the effort centered on the design of the network interface. This design was shaped by Davies's assumption that businesspeople w~uldturn out to be the main users of networks. Davies (l966b, p. 3) explained: "The emphasis on real-time business systems in this report is due to the belief that they will generate more real-time digital communication traffic than, say, scientific calcuJations or computer­ aided design." Whereas academic researchers might need to transfer large amounts of data from one computer to another, businesspeople would be using terminals to access an interactive computer. The NPL designers therefore focused mainly on providing an easy-to-use termi­ nal interface to the network. 33 32 Chapter 1

One unusual characteristic of the Mark I that derived from the emphasis on user friendliness was that all terminals, printers, and other peripheral devices were connected directly to the network. The network was actually interposed between a computer and its own peripherals, so that .the network became, in a sense, internal to the computer. Davies (1966b, p. 11) commented: The overall description of the system shows a major organisational change. Present day multi-access computers each have equipment which assembles messages from keyboards and distributes them to printers. Whatwe9're proposing is that this function should be carried out by the network, not the attached computers.

Using the network as a common communication channel ft;u'all corn­ ponents would make it possible for apy pair of machines to interact, Normally, a terminal user who wanted to print a file would have to log in to a host computer and send a command to a printer attached to that computer. With the Mark 1, howfver, the user could send a command directly from their terminal to the' printer, without ever having to go through the printer's host computer. Remote resources would be as easy to use as local ones, since the access procedures were If.~~y.r~1.4 .. .' . " ...... identical. This was a radical concept in user interface design-a con­ AMark I terminal. The text pn the screen reads NPL Data COnll11Ul1lCatlons cept that would nor become a' commonplace of networked Network." Source: National Physical Laboratory, Teddington. Reproduced by systems for another twenty years. permission of controller of HMSo. There was a price'to pay for this visi?n, however. Since all terminals

were connected through the network, a failure in the network would 1'~Impact o!Da,vies's Work . . mean that terminals would be cut off even from their local host com­ The Mark I carne to be used regularly by researchers at the National purer." The variety of peripherals attached to the network also made Physical Laboratory, and in 1973 Donald Davies's team introduced an the interface computer more complicated and expensive to build, upgraded version of the system called "Mark n." The. Mark n used which delayed the completion of the project.27 And, in trying to make most of the same hardware as the Mark I, but software unprovements the terminal interface user friendly, the designers of the Mark I sac­ made it two to three times faster. The Mark n remained in service at rificed flexibility and adaptability. For example, they implemented the NPL until 19864uitean impressive term of sen/ice for an experi­ parts of the user interface in hardware (figure 1.4). A user wishing to mental systerp (Campbell-Kelly 1988, pp. 237~239).Drawing on their

set up a connection would punch a button marked TRANSMIT on experience with the network, members of thf NPL .tea~nwe~ton to the front of the terminal, after which a light labeled SEND would light participate in several larger network projrcts m the United Kingdom up to indicate that the network was ready to accept data; there were and in Europe. other lights and buttons for different operations. This interface was But despite Davies's technical innovations and the local success of easy for novices to learn, but it was harder to automate or modify than the system, the Mark I did not have the kind ofin~ueucetha~the a procedure implemented in software would have been (anonymous ARPANJ<:T.would have. Davies. was never able. to build the national 1967). In the fast-changing world of computing, a system that was not network he had proposed, and the specific techniques used in the adaptable was in danger of becoming obsolete. Mark I were not transferred the NPL Though Davies had 34 Chapter 1 White Heat and Cold War 35

a head start on the builders of the ARPANET, it was their work that the GPO's reluctance.to nrovide

would come to dominate the field of computer ne.tworking: ...... •.

The pplitics of t!le day alld the C41turl;:·of sqme British institu~ions hampered Pavies's ability to implement hi . his ailjl of entrepreneurs discovered they were all hampered in ~heirtime sharing ....·;l,cti,litil=sby the same thing-wh~tthey felt was foot-dr~ggmg0::the P~l~tof keeping the l)~jtedRingqom ahead of 1Il cQplpl.lter •• t~';.:·r,1>0 ... when it came to lines and modems for ume shallng set vrees. networking. In t4elatel950s the NPLhad een oriented towarclpure p.370)28 research, but under the Wilson government there was a marked in­ crease in government oversight and intervention. In ~herecollection members decided that public action was called t01~and on.3 July of one NPL scientist (Pyatt 19.83, pp. 145:-146): 8 they held a public event ~alled"Conversa:tiopal Computmg on South Bank" at London's. Roy~Fe~tivalHall (Campbe~-Kelly Schemes for improving the service given to the 228). Commercia) time sharillg firms demonstrated their serv- hawked from above.... Open-ended research was seve its place all research projects had to have a 'customer,' who leading figures in computing talks, and hun~redsof com- suaded of the viability and value project and agree to make available .: professionals attended. One the club's leadmg m~mbers, the funds to carry itouc [witlI customers~required regular y Gill, a professor at Imperial College, gave a speech ~rgmgth~t preparation of cases by in time which could ill be spared d Davies's network design be adopted. The Amencans~Gill from practical wode ,. were already working on plans for the ~PANET:~ell For Davies and the Mark I team, the emphasis on promoting CO and widely reported the pr~ss,Conversatt~llalC~m~ut~ng South Bank generated a public der

Putting It All Together: Packet Switching and the ARPANET research in universities. In

Paul Baran and Donald Davies had both envisioned nationwide net­ variQus,federal agencies, said that this mom:y should to works that would use the new technique ofpac~et~~itching,but lish centers .of excellence:' throughout the natioa (Johnson neither mgn had been able to fully realize thiagoal, Instead, the first He urged each government agency engaged in research large-scale packet switching network would be built by the Advanced to take " all practicalmeasmes. . to institutions where Research Projects Agency.32The design of this network would draw 011 res!ealcch now goes· on,· and to help additionalinstitutions to become the work of both Baran and Davies,·but the network's builders had more effectivecenters.for teaching and (ibid., p. 336). their own vision of what packet switfhing could achieve. johnson specifically did .not.want to limit these centers to ARPA was one of many new American science and technology ven­ missien-osaeneed projects. policy," he wrote (ibid.,p. 335), tures that had been prompted by the Cold War, Founded ill 1958 in lIl()re:sllPl:lO]l'twill be prov~dedunder terms which give the university response to Sputnik, ARPA had as its stated' mission keeping the and the investigator wider scope for inqu~ry,as contrasted with highly United States ahead of its military rivals by pursuingresearch projects sp.e.cific. narrowly definedprojects," 33 that promise significant advances in defense-related fields. Through­ few!J;lon!bs Iaten. the. Department of Defense responded ..to out its existence ARPA has remained a small agency with no laborato­ jQhnson's call w~tha plan to create ries of its own. ARPA managers initiate and manage projects, but the related research. +.'Each new univt:rsilty pr'Q~Jl';.u:ll,.'!tble actual research and development is dope by academic and industry '\should stimulating and, contractors. Recognized even by its critics for good management and at;the same time, .•·contribute to for.solving rapid development of new technologies, ARPA has had some success problems in national defense.";(D..epartmentofDefense.19'72,p. 33'7) in transferring its technologies to the armed and the private IPTO created several computing research centers, giving large. grants sector (Pollack 1989, p. 8). toMn~Carnegie Mellon, UCLA, and other universities. By 19'70, The director of ARPA reports to the .Directo};ofPefense Res.earcl.1 ARPA had .funded a variety of time sharing. computers located at and Engineering at the Office of the SecretElry pf Befense. ARPA has universities and other computing research sites across the United several project offi.f:es that fund research in differ·ent areas; project ~tates.The purpose of its proposed networ!,(-the ARPANET-was to

chang~ng offices are created Or disbap4ed in response to the priorities <.:0l1nectthe~esc::atfereel cP~PUti11gsi~Fs. of the Department of Defense. Each office has a director and several TIlI;.i\RPANETproject ;Wa~maJ;l~gedbya com- program managers, all of whom are directly involved in choosing puterscientist who had conductesl networkingexperirnelltS at NUT's research projects. The first project offices directed research in behav~ Lincoln Laboratory before joining ARPA in 1966. Roberts. had a man­ ioral sciences, materials sciences, and missile defense. In 1962,with elate to build e lergF' multi-colnputernetworkr .put he did. not initially the founding of its Information Processing TechniqUeS Office (lpTO), have a finn idea of how to do this. l:Ie'.considered having pairs of ARPAbecame a major funder of computerscience inthe United States; computers establish a <.:onnection using-ordinary telephonFcalls when- often outspending universities significantly. Compllfer science, not yet they needed to 1xchange d~tap-a p1ethQd l1ehad employed in an established discipline in 1962, developed rapidly once IpTO began earlier experiments. But the high cost of long-distance telephone con­ funding it. IPTO has been the driving force behind several important nections made tlli.~option seem prohibitively expensive. Roberts also areas of computing research in the United States, including graphics, worried that 0rqinarx phone service woulq he una<.:ceptably prone to artificial intelligence, time sharing operatingsystems, and networking tra.ns~~ssionerrors and ,line failures, AlthOllgh he was •~wareof the (Norberg and O'Neill 1996).34 concept ofpac~1te"(ftching,Roperts was 110t sllre to implement ARPA's funding of basic tesearch was consistent with the philosophy it i11alargellF~"(ork... . of the administration of President Lyndon johnson, who, in a Septem­ IJ;l October..ofJ967, uaresorved, ber 1965 memo to his cabinet, advocated the use of agency fUhds to attended a computing svmoosium 38 Chapter 1 White Heat and Cold War 39

was slated to present ARPA's tentative networking plans. Roger Scan­ of the ARPAI\lET contractors, tlebury of Britain's National Physical Laboratory also presented a Leonard Kleinrock, were used it paper at the symposium, where Roberts heard for the first time about in their research.57 In 1967, Roberts recruited Baran to advise the Davies's ideas on packet switching and the ongoing work on the A.RPAt"JETplanning group on distributed communications and packet Mark 1. After this session, a number of conference attendees gathered switching. to discuss network design informally, and Scantlebury and his col­ ;Through these various encounters, Roberts and others members of leagues advocated packet switching as a solution to Roberts's concerns the ARPANET group were exposed to the ideas and techniques of about line efficiency. The NPL group influenced a number of Ameli­ Baran and Davies, and they became convinced that packet switching can computer scientists in favor of the new technique, and they and distributed networking would be both feasible and desirable for adopted Davies's term "packet switching" to refer to this type of net­ the ARPANET Packet switching promised to make mere efficient use work. Roberts also adopted some specific aspects of the N~Ldesign. the network's long-distance communications links and to enhance For instance, Roberts had planned to use relatively low-speed tele­ the system's ability to recover from equipment failures, which an phone lines to connect the network nodes. He later recalled that, after experimental network would surely encounter. At the same time, how­ the NPL representatives had "spent all night with [him] arguing about ever, packet switching was an unproven technique that would be the thing back and forth," he had "concluded from those arguments difficult to implement successfully. The decision/to employ packet that wider bandwidths would be useful" (Roberts 1989). Roberts switching on such a large scale reflected ARPA's commitment to high­ decided to increase the bandwidth of the links in his proposed network risk research: if it worked, the. payoff would be .not. only greater from 9.6 to 56 kilobits per second. The ARPANET would also use a efficiency and ruggedness intheARPANETitsdf, butalsQa significant packet format similar to the NPL Mark 1.'15 advance in computer scientists' understanding properties After the ARPANET project was underway, the acoustics and com­ and techniques. The ARPA managers could, afford (indeed, had a puting firm of Bolt, Beranek and Newman, which had the main mandate) to think extravagantly-to aim for the highest payoff rather contract to build the network nodes, continued to interact with. the than the safestinvestment. NPL group. According to BBN's Robert Kahn (1990), The Social Construction of Packet Switching Donald Davies was a very creative guy; he thought a lot about interesting ideas of how networks should be built. He clearly had the concept in his head of what packet networks ought to look like, and he had done it independently The projects sponsored by Rand, the NPL, and ARPA had much in in England. I believe Larry Roberts will probably tell you that Donald had a common in their approach to packet switching, but some crucial dif­ big influence on him. ferences in ARPA's approach helped the ..i\RPANET play a more enduring and influential role than the other.proj~cts.DonaldDavies, The NPL's Derek Barber visited the BBN team in 1969; he reported Paul Baran, and Lawrence Roberts each made technical choices based that they "were interested in the possibility of connecting our type of on specifieloeal.concems, and the extent to which their systems were local area [network] directly into" the ARPANET and that they saw the influential depended in part on whether others shared those concerns. NPL work as "complementary" to the ARPANET project (Barber 1969, p. 15).36 For instance, Baran's system had.maay.elements that were specifically adapted to the Cold War. threat, including very high levels of redun­ Paul Baran, too, became directly involved in the early stages of dancy, location of nodes away from population centers, integration planning the ARPANET Roger Scantlebury had referred Lawrence and of cryptographic capabilities and priority/precedenc~features into the Roberts to Baran's earlier work. Soon after returning to Washington system's design. None·of these features were adoptediby Davies.or from Gatlinburg, Roberts had read Baran's On Distributed Communica­ Roberts, neither of whom was concerned with survivahility.38 On the tions. Later he would describe this as a kind of revelation:' "Suddenly other hand, ;aspects of Baran's system that would be useful in a variety I learned how to route packets." (Norberg and O'Neill 1996, p. 166) 40 Chapter 1 Heat War 41 of situations-such as high-speed transmission, adaptive routing, and efficient packet switching-s-were adopted for use in later svstems. Wn,hprt,,>h:lVP a he was One thing that Baran, Davies, and Roberts had in common was the computer experts from around the country to help build the insight that the capabilities of a new generation of small but fast network. Davies, at the NPL, had a much smaller budget. Faced with computers could be harnessed to transcend the limitations of previous a perceived economic crisis and convinced of the need to with communications systems ..Telephone systems ofthelate 1960s did not the United States and other of high technology, the British use computerized switches, and message switchillg systems used large, government tried to rationalize the computing industry and to encour- expensive computers that handled messages slowly. Whewpresented age commercial spinoffs of government Eventually much of with the idea that a network could employ dozens of computers as its the research at the NPL and at similar was focused on switches, people in the communications industry tended todoubt that short-term commercial applications, and the Labour government's computers fast and cheap enough to make this idea feasible would be industrial policy limited Davies's choice of computers. The US govern­ available (Baran 1990, pp. 19-21; Roberts 1978, p. 1307; Roberts inclined to try to manage the domestic computer indus­ 1988, p. 150; Campbell-Kelly 1988, P: 8). Indeed, the first of these try. Overall, Roberts had much more support and much less small but powerful "minicomputers" did not appear until 1965, when interference from his government than Davies had from his. the Digital Equipment Corporation introduced its PDP-8. The fact that Davies had been one of the earliest and most articulate advocates of packet switching relied on an innovative computer product helps to packet switching. He had a national explain why that technique was consistently explored by computer network at a time when the ARPAl'N"ET Yet the scientists but not by communications experts, even though it drew on middle of 1968 Davies was project had been aspects of both fields. eclipsed by the American effort: "As a force in this discussion NPL is In the 1960s, computing technologies became policy instruments too remote and our own demonstration as planned now is small-scale both in the United States and in the Unite~Kingdom. In the United and likely to be delayed by the reductions ~n and administrative Kingdom, intervention in the computer industry was seen as a symbol difficulties in purchasing computers." (D

at

a

the

the

the

early

com­

data,

com­

inter­

Lick-

serve

and

Taylor,

devices,

of

comput­

an

to eliminate

director

reasoned:

interastive

permit

became

many

funded

The

In

in

first

Robert

«:1,,4

(IPTO).

To

output

industry

would

including

practice

retrieving

need.

step

already

interactive

a

.and

the

Licklider

users.

for

sharing.

Office

in

next

and

of

network

made

and

input

(1970, p. 543)

creative

systems

time

the

IPTO,

communities."

macmnes

"symbiosis,"

user,

the

of

of

most

better

which

methods

like

changes

Challeng~s

in

such

C. R.

numbers

effective

about

sharing

Wessler

the

inspiration

on

human

An

an

LiSflider's

director

"Man-Computer

Lickador

large

specific

systems, and seemed

time

the

extension

intuitive

languages,

engineer

for

computer of

tobring

from

paper·

together

heeded

limitations

intellectually

inifonmatioll1-l1landliJIIK

identify

sharing

resources.

back

more.

logical.

Roberts

machine.

born

be

ARPANET: to

human

feasible

the

a

needed

1960

in

who

networking

community,

a systems

time

the

was

on

by

coupled

aspirations

of reorienting

distance

traced

to

as no

paper,

and

programming

that,

be

were

should

software

those

been

Information

is

for

and

communications.

went

a local

mankind.

of

and

will

think

computers,

that

adapt

scientists,

of

influential

years

had

can

data

size

economically

to

approached

needs

Among

ARPA's

the

sharing

"W~thin

ARPANET

computing

~l.lter

inl;{ development

and

hi~her-Ievel active

pU#Bg

Lick}ider

machines

user

IIl,:l;mte1,to

of

~t,.ategies

~uildingthe 2 44 Chapter 2 45

NASA before joining ARPA in ]965. 1 By the mid 1960s IPTO was funding computing research centers around the country to work on projects such as time sharing, artificial intelligence, and gracphi~;s. Taylor (I 989) felt that each ofIPTO's scattered research centers "had its own sense of community and was digitally isolated from the other one." When he became director of IPTO, in ]966, Taylor began to speculate on ways to "build metacommunities out of these by connect­ ing them" (ibid.). Early that year; he and ARPAdirector Charles Herz­ feld discussed a plan to link IPTO's computing sites with an experimental network. In 1967 Herzfeld agreed to allocate $500,000 for preliminary work on the idea, which was dubbed the ARPA Net­ work Or ARPANET: Figures 2.1 and 2.2 illustrate how the AJ,{PANET reproduced the geography of ARPA'srese.an::b network, spanning the United States to link ARPA'S computing' sites. Besides serving Taylor'svision of linking the research comruunity, the network would address a pressing need within ARPA. ARPA was creation. the major funding sourcefor most ofits computing' contractors, and buying computers for them represented a large expense, for the agency. To make matters worse, a single contractor might need access to several types of machines. Computer hardwar~and qperatinp-. sys­ tems tended to be optimized for PilniH1lilr lls(;~,sl,l~has in..terilctive time sharing or high-Powere~l"n4mger<;t:unI;Ping.': •COmPuters .also had a variety of specialized-input/outpuc devices, such as graphics terminals. Contractors who wanted to combine different modes of computing had to either travel another site or acquire multiple machines. Asa result, IPTO wils continually besieged by requests from its contractors for more computers. Taylor believed that if ARPA's scattered computers could be linked together,han:Jware, software, and data could be efficiently pooled among contractors rather than waste­ fully duplicated. In late 1966 Taylor recruited Lawrence Roberts, il program man­ ager at MIT's Lincoln Laboratory, to oveneedevelopment of the ARPANET: Roberts had been pursuing networking experiments at the Figure 2.2 A lllap of the fifteell-nqde AFlPl\N,ET redrawn from Bolt, Beranek Lincoln Lab, and Taylor considered him the best qll~lifiedcandidate and..Newll1\.in's qriginal. (SPC;: SystenlS D1evl:lOI:>Q1ent C.orporatiop.CMLJ: to manage the ARPANET project, but J,{ogens \Vflsinitially reluctant Carnegie Mellon University) to leave his research position. The circumstances of his joining IPTO provide an example of ARPA's leverage over the compllter science research community. When Roberts turned down an initial invitation to come to ARPA, Taylor asked ARPA's director, Charles Herzfeld, to call the head of the Lincoln Lab and remind him that half of his lab's funding came from ARPA, and that it would be in the lab's best 46 Chapter 2 ARR4NET 47 interests to send Roberts to Washington.' Roberts joined lITO as cOirnrnunic:auons, To the project on Roberts denloved assistant director and became director when Taylor left the agency in a.uninue set of technical and managerial Both the ARPA- March of 1969. itself and ARPA's approach to building it would have a lasting Roberts envisioned the ARPANET as a way to bring researchers influenl:e on the emerging field together. He stressed early on that "a network would foster the 'com­ munity' use of computers." "Cooperative programming," he contin­ InitialChaUenges ued, "would be stimulated, and in particular fields or disciplines it will be possible to achieve a.'critical mass' of talent by allowing geographi­ .J,

Another unusual and potentially troublesome characteristic of the published the fall of 1 ARPANET was the great variety of computers it would connect. Roberts left Lincoln Lab for ARPA. Besides machines commercially available from IBM, DEC, GE, SDS, In describing their experiment, Marill and Roberts articulated some and UNIVAC, the proposed ARPANET sites had various one-of-a-kind important concepts. In their view, the "elementary approach" to con­ machines, such as ARpA's experimental ILLIAC supercomputer (Dick­ pecting two computers was for each computer to treat the other as a son 1968, p. 132). These various types of computers were incompatible tel·milna.l. Such a connection required little modification of the com­ with one another, which meant that users who wanted access to pro­ puters, but it had severe limitations. The connection was slow, since grams or data at other sites often had to reprogram the. software or terminals operate at much lower data rates than computers, and there reformat the data. In 1969, ARPA director Eberhardt Rechtin told was no general-purpose way to access a remote system, since each Congress: "When one user wants to take advantage of another's separate application program had to manage its own connections developments, he presently has little recourse except to buy an appro­ rather than having the operating system handle the connections for priate machine or to convert all of the original software to his own all applications. Marill and Roberts thought that forgoing the elernen- machines." (US Congress 1969, p. 809) Incompatibility wasted time approach and taking on the harder task. of modifying the com­ and programming resources, and it remained an obstacle to collabo­ PIJters' operating systems would make it possible to create a rative work. For Roberts, one aim of the ARPANET project was to h~gher-speedcomputer-to-computer interface instead of relying on overcome these obstacles. Roberts viewed the diversity of computers the ordinary tenninal.,w-computer interface. They proposed that each not as an unfortunate necessity but as a strength. of the system, since Qost computer implement a general-pIJrpose set of rules for handling a network that connected heterogeneous systems could offer users a 4 network, connecrion, wh~chtheycalledthe "meSS4ge protocol" (Marill wider range of resources. But getting this assortment of machines to and Roberts 1966, p.428). had learned from communicate would require an enormous effort in hardware and this experinJ.ent to the. design. of.the decided that all software redesign. "Almost every conceivable .itemof.computer hard­ the host computers should follow a stand4rd protocol for network ware and software will be in the network," Roberts pointed out, adding interactions.. liaving a standard protocol would help overcome the "This is the greatest challenge of the system, as well as its greatest in<;ompatibilitiespetween different types of computers.. However, this ultimate value." (quoted in Dickson 1968, p. 131) aPproach also created a huge: task for the people maintaining the Roberts's view was based on his experience as one of the first people hosts, who would have to add th,is new networking capability to the to attempt to establish a connection between different types of com­ operating systems of their various comptlters. puters. After receiving his Ph.D. in Electrical Engineering from .MIT Creating a heterogeneous, packet switching, continent-spanning in 1959, Roberts began working at the Lincoln Laboratory, where he computer-to-computer network would be a significant technical became interested in the possibility of networking computers for time achievement for ARPA; the challengewotlld lie in keeping these. same sharing during discussions with]. C. R. Licklider, Donald Davies, and t~aturesfrom leading the project into Ch40S. The technical and.mana­ others in 1964 and 1965 (Roberts 1988, pp. 143~144).Roberts found gerial difficulties of the ARPANET project became apparent when a kindred spirit in Thomas Marill, who had studied under Licklider Taylor and Roberts presented the network concept at IPTO's annual and had founded a time sharing company in Camhridge called the meeting of Principal Investigators (scientists. heading researcl; pro­ Computer Corporation of America, In 1966, with funding from IPTO, jects) at the University of Michigan in April of 1967. Roberts had Roberts and Marill undertook to build a rudimentary network linking already discussed the idea infonpally with several of the PIs, but at the two experimental computers: the TX-2 at the Lincoln Lab and the meeting he announced that the project would definitely go forward. Q-32 at the System Development Corporation in Santa Monica. A line The PIs,1Vho would have to design, iQ.lplement, and use the proposed leased from Western Union provided the communications link, and network, did not gr.eet the net1Vol?kidea the enthusiasm it would Marill and Roberts wrote their own software to manage the connec- receive in later years. Most 50 Chapter2 ARPANET 51

or even hostility to the idea of connecting their computer centers to p.la,p.agt;Ill'Ent style. approach

the network. Some of them suspected-correctly-that ARPA saw the iC~Q.im

~e(:hruqjllesan influence beyond their-role as management tools for the Although they knew in the back of their mind that it was a good idea and were supportive on a philosophical front, from a practical point of view, they-Minsky, and Mcflarthv," and everybody with their own machine-s­ wanted [to continue having] their own machine. It was only a couple years after they had gotten on [the ARPANET] that they staneci raving about hCM lavered system is organized as a set of discrete functions that interact they could now share research, and jointly publish papers, and do other things 4~cordingto specified rules. The functions are called.tlayers" because that they could never do before. are arranged in a conceptual.hierarchy that proceeds from the Many PIs did not want to lose control of their local computers to concrete and physical functions (such as handling electrical sig­ people at other sites, and they saw the network as anintrusion.t Since most abstract functions (e.g., interpreting human-language "their" machines were actually paid for by ARPA, the PIs had little users). Each higher-level function builds on the capa­ choice in the matter; however, they were not eager to join in the provided by the layers below. The idea of layering seems to network. Even those who agreed on the general advantages of devel­ occurred independently to many people working on networks as oping computer networks had practical objections to implementing drew on concepts of modularity and functional division of systems the ambitious system envisioned by Roberts and Taylor. Some of these were current in.compuser :Sclence,6 PIs were unwilling to undertake the massive effort that seemed to be the ideal layered system, theppp(';)rtunities interaction among required; others were convinced that the project would fail altogether. limited and follow Tllis reduces the complexity of Besides reminding us that even those at the forefront of computer system, making it easier to design, test, and debug. The designer science in 1967 could not foresee the astounding popularity of-the particular layer needs taknow how that layer is expected to ARPANET and its successors, the negative reactionsofthe Principal int:er;act with other layers does not need to know anything about Investigators illustrate the two majorchaUenges that ARPAfaced. First, internal workings of those layers. Since the layers are independent, it was dear that the complexity of the network's design would require can be created and modified separately as long as all those imaginative technical solutions, Second, ARPAwould need to find ways vvorking on the system agree to use the same layers. to gain the cooperation of prospective network members. The PIs were Thus, layering has both and social it makes the initially more concerned with continuing their own local projects than t(;chnical complexity of and it allows the with collaborating on a network. In order for the project to succeed, system to be designed way. Lawrence Roberts would need to create some sense of common The ARPANET's a specific plan for purpose. how functions would be how the interfaces and protocols would work. evolved as the System- Building Strategies AJ~PANEIdeveloped. The layered approach was taken at the 1967 nleetillg Investiga,tors. One of the Of the many problem-solving strategies that Roberts and his team of contractors' main concerns on first hearing about the project was that contractors would employ in building the ARPANET, two were espe­ creating the necessary packet switching software for their computers cially significant. One was an approach that came to be known as would require too much effort on their part. IPTO research sites used lasering, which involved dividing complex networking tasks into modu­ a wide variety of time sharing operating systems; if the host computers lar building blocks. The second was an informal and decentralized had to perform packet switching, someone would have to nrn

each different type of computer to perform the various packet switch­ 2.1 ing tasks and then reprogram each computer whenever the software two-layer model of the ARPANET. needed modification. Moreover, the packet switching software would have to be designed to accommodate the limitations and idiosyncrasies Functions of each model of computer. In view of these difficulties, even PIs who Handles user interface; initiates and maintains were sympathetic to the project's goals had reason to be skeptical about connections hosts its technical feasibility. Moves data through saoaerusme One of the Principal Investigators, Wesley Clark of Washington hc)stcilVll.l'and University in St. Louis, saw an easier alternative. Clark was familiar with the capabilities of minicomputers, and after the meeting he sug­ gested to Roberts that each of the host computers be attached to a special minicomputer that would act as the host's interface to the one pan of the network to another, and the network. In Clark's plan, the minicomputers, rather than the hosts, r~SPl')n~Sib:lefor the content would form the nodes of the network and handle the packet switching coulclnow be wrinen operations. This network of minicomputers was designated the clitrerent types of Since minicomputers were becoming relatively inexpensive by the late a "black box" 1960s, it seemed economically feasible to dedicate several of them to how it worked, running the network. Taylor endorsed the subnet scheme, and Roberts incorporated it into the ARPANET design's, calling the mini­ computers "interface message processors" (IMPs).7 Figure 2.3 illus­ trates the subnet idea. The subnet design created a division of labor between the switching nodes (IMPs), whose task was to move packets. efficiently and reliably of layers or protocols, the two-laver model (table.2. suggests . kinds of relations between functions. First, the flinction-s beCOme increasingly abstract as from the b()ttQm tothe.tqp of the stack__from moving

OVer wires to int~rpretipg.co,mmap?s. tYP7cl Py terminal Second, the qrd7f oftht: layers J:ePr~sentsatt:mJ?oral sequence top to bottorn: first the.·ps7r type~

::::Ieased "protocol stack" model would 911ickly come t~dominate th.e telephone line people •though~about organizing networ~spreCIsely because It ! offered a blueprint for reduc~ugthe q~mplexltyof network compo- while increasing the PI7dictabHity of the system a whole." the ARPANET was finish7d the model would be ex.p~ndedto Figure 2.3 layers, and in- later . still would be added Network model with communications subnet. with and how to on~apjze 55 54 Chapter 2

Informal Management Whereas the layering approach stressed separating the system's ele­ people who were receiving••.s1Jlpport ments, ARPA's management style was aimed at fostering the coopera­ a special point of providing ongoing funqing for gradu­ tion required to integrate those elements into a coherent whole. stu.dents at. contract sites, and he arranged special meetings.and ARPA's unmatched financial resources drew many computer scientists ~;orkinggroups for them (Taylor 1989, p. 19). Graduates of the IPTO­ into its projects, but ARPA managers did not conduct relations with flJln¢led programs at MIT, Stanford, Carnegie Mellon,.and elsewhere their researchers on a purely financial, contractual basis. The organ­ major source of computer science faculty atAnlerican uni- izational culture surrounding the ARPANET was notably decentral­ ve~;sities,meretw extending ARPA's network into the next ized, collegial, and informal. In coordinating its contractors, ARPA er;:lt~O!nof researchers (Norberg and O:Neill 1996, pp.290-291Vo relied largely on collaborative q.rrangements rather than contractual of the expense ofc()mputing machinery in the 1960s and obligations, and technical decisions were usually made by consensus. large role in funding[computer science, IPTO managers had The network itself provided a new way to coordinate dispersed activi­ real power over their contractors, and they were willing to use this ties and came to function as a meeting place for the computer science power when they felt it necessary. As has already been noted, Robert community. Though conflicts sometimes arose among the contractors, Taylor exerted pressure on Lawrence Roberts to leave his position at the ARPANET culture enhanced ARPA's ability to enlist the support the .Lincoln Labandjoin ARPA. Once in. charge of the project, Roberts of the research community and to respond to the technical challenges . hesitate to make reluctant contractors share in the ARPANET that the project posed. The collegial management style of Taylor and Roberts was typical universities were being funded by uS'flpd we said, "We are going to build of IPTO in the 1960s and the 1970s. IPTO recruited most of its a'network and you are going topartid~a~ein it. And you are goin~to Connect directors and project managers from the ranks of active researchers at your machines. By virtue ofthat.weare going to reduce our compunng university and industrial research centers. IPTO managers kept in demallds on the pffice. So that going to buy you touch with their colleagues by touring contract sites to evaluate the comnuters untilyou have l~sedupJlll the network." progress of programs, learn about new ideas, and recruit promising time we started forcing them. Wbe 1989) researchers. Not career managers, they generally stayed at ARPA only But IPTO managers preferred to take the informal approach when­ a few years before returning to academia or private business (in part ever possible. Having been researchers themselves, they subscribed to because ARPAsalaries were modest). Though ARPAas an organization the view that the best way to get results in basic research was to find had financial power over its contractors, most of the individuals who talented people and give them room to work as they saw fit. They also actually managed IPTO projects were drawn from those contractors. tellded to believe that differences of opinion could be debated ration­ Howard Frank of the Network Analysis Corporation, an ARPANET ally by the parties involved and decided on their technical merits, and contractor, observed: "It's easy to say 'the government,' or ARPA, or that they, as lPTO managers, would need'm intervene with an execu­ something like that, but they are individuals that you deal with." tive decision only if the contractors could not resolve differences (Frank 1990, p. 300) among themselves. Not surprisingly, J,PTO contractors praised this ARPA tended to award contracts through an informal process, fund­ management style as an enlightened and productive way to conduct ing individuals or organizations who were already known to IPTO research. The report of an outside consultant commissioned by ARPA managers for their expertise in a particular area. The ARPAapproach to report on the project's status in 1972 agreed that the project's exhibited the weaknesses and the advantages of an "old boy" network. informal style had contributed to its 'success, nocing'that the process Many talented computer scientists found themselves left out of the ofbuilding··.the ARPANET had "been handled in a rather informal field's biggest funding opportunity, but those who were included en­ fushicm with a great deal ofauton'on1y and anindeflnite division of joyed an extremely supportive environment." Wesley Clark (1990) commented: "In the ARPA system, once you were in, you were a 56 Chapter2 57 telephone conversations, and understandings are relied upon for day main development was to to day operation. This environment is a natural outcome of the pro­ projects, which were initiated gressive R&D atmosphere that was necessary for the development and , ,,~(JWnint~~restand expertise in a implementation of the network concept." (RCA Service Company ." j,ng. C(~mputerwas a new venmse initi<~te'~ 1972, p. 34)11 can.didate the job.

In view of the nature of the project, it made sense for Roberts to m~'.. 1> usual practice encourage ARPA's network contractors to work together as peers. from a numher of computer and engilt1el:ring Different tasks required different combinations of skills, and no one after considering bids from a dozen contractor had the overall expertise or authority to direct the others awarded the contract to the Bolt, .ti(~raneKand Newman as subordinates. Roberts's informal coordination methods provided a ~:.orpoJratllonof Cambridge, a re!;:ttn'ely context in which the network builders could, for the most part, share spedaJlIZlTlg in and CQrnplltin,g S)fstems. skills and insights on an equal and cordial basis. Though 11.0ta giant in the co:rpp'utl:r business, Bolt, and ~ewnlanhad several company had Getting Started ,pre,yiolllS ues,:\IV'ltll1PTO: J. C. R. at BaN I,elore Roberts began the ARPANET project informally. Rather than soliciting IPTO's bids for contracts right away, he brought together a small group of had Principal Investigators who had expressed interest in the network 1>411'<::;:' in the concept and began meeting with them to discuss design problems and . f:lLan.n1r1gof the ARPANET., BBN was to work out possible solutions. He asked Elmer Shapiro of the Stanford In 1990, Kahn Research Institute to lead these meetings, and he recruited Paul Baran • ,pyl;>ri,d version of Harvard and of the Rand Corporation (the man who had done the earliest work on peiopile there were either facull:y JorfOI7merf~l:ultyote!lherHLos Angeles. In June of 1968, Roberts actual use, easily programmed, and equipped with good .input/output submitted the plan his group had worked out to ARPA director Herz­ capabilities (Heart et al. 1970, p. 557). BBN and Honeywell were feld. In July, Roberts received an initial development budgt;t of $2.2 located conveniently near each other in the Boston area, and the two million and approval to seek contractors to dt;velop the nerwonk.v cOlnpanies had agreed to work together on customizing the H-516 for The basic infrastructure of the ARPANET would consist of time llse in the network if BBN's bid were accepted. sharing hosts, packet switching interface message processors,and Other 1"0lltracts were awarded less forl1lally. Since the network was leased 56-kilobits-per-second telephone lines to connect the IMPs. The supposed to.bea experiment as as a cOlmnllUJllil;::ati~ons hosts were already in place, and the lines would be prqvided by AT&T; Roberts awarded a contract t.o to ARPAl,lET 59 58 Chapter 2 create theoretical models of the network and to analyze its actual running the of performance. Roberts knew Kleinrock from MIT, where they had been pologies and constraints, NAC was able to provide a range of solutions graduate students together, and he was aware that Kleinrock had that satisfied ARPA's performance criteria while minimizing costs adapted a mathematical tool called queuing theory to analyze network (Frank, Frisch, and Chou 1970, p.S81), systems. UCLA was designated the Network Measurement Center, and Roberts gave the Stanford Research Institute a contract to create an Kleinrock and his students would take an active part in testing and online resource called the Network Information Center, which would refining the network. To allow the analysis to begin as soon as possible, :maintain a directory of the network personnel .at each site.. create an UCLA was chosen as the site of the first IMP online archive of documents relating to the network, and provide Once the initial four-node network was functioning smoothly, information on resources available through the network (Ornstein Roberts planned to extend the ARPANET to fifteen computer science et al. 1972, p. 253). In addition to Elmer Shapiro and a number of sites funded by IPTO (figure 2.2), then to additional ARPA research other computer scientists who were interested in the network, SRI had centers doing work in other fields, and then perhaps to military sites on its staff Douglas Engelhart, a pioneer of human-computer interface (Roberts and Wessler 1970, p. 548). With fifteen or more sites, it would design who, in 1965, had invented the mouse. Engelbart was develop­ be no simple matter to design a network that combined the redundant ing a database, a text-preparation system, and a messaging system with connections needed for reliability with maximum data throughput and a sophisticated, user-friendly interface. Roberts hoped that, with such minimum cost. Thus, Roberts contracted the Network Analysis Cor­ tools available, SRI would provide an easy-to-use information center poration (headed by Howard Frank) to help in planning the network's for the network community. topology-the layout of nodes and communications links. Frank had Roberts also reestablished his informal networking group, now met Kleinrock when they were both lecturing at Berkeley, and Klein­ named the Network Working Group (NWG), to develop software rock had later introduced Frank to Roberts. Frank, whose experience specifications for the host computers •and to provide a forum for included optimizing the layout of oil pipelines, had formed NAC to discussing early experiences and experiments with the network. The provide consulting services for businesses building complex systems. most active members of this group were computer science graduate For Frank, taking part in ARPA's cutting-edge research project was a students who had been asked by their advisers to represent their sites. welcome change from more routine commercial work, and it was also At UCLA, which was particularly active in the NWG, Leonard Klein­ a way to keep his employees motivated: "It wasthe more intellectually rock was using ARPA money to support a number of Ph.D. students, challenging of the work. And I could have really smart, capable people including Stephen Crocker, Vinton Cerf, and Jon Postel, all of whom working on that ... where, if you only put them into commercial were to be important figures in the development of host software. applications, after a while they would just leave. They would get very Other active members in the early stages of theNWG included Jeff burned out." (Frank 1990) Frank's team would use recently developed Rulifson, Bill English, and Bill Duvall at SRI, Gerard Deloche at computer simulation techniques to evaluate the cost and performance UClA, Steve Carr at the University of Utah, Ron Stoughton at UC of various network topologies. NAC's analytical tool was a "heuristic" Santa Barbara, John Haefner at Rand, Bob Kahn and Dave Walden computer program-one that provides an approximate solution to a at Bolt, Beranek and Newman, and AbhayBhushan at MIT:13 The problem whose exact solution would require too much computing membership of the group changed constantly, growing larger as more time. They would give as input to the program an initial topology that sites got involved in the network. The NWG would gradually develop satisfied the performance constraints specified by ARPA: a maximum a culture of its own and a style of approaching problems based on the delay of 0.2 second for message delivery (for responsiveness), a mini­ needs and interests of its members. mum of two links per IMP (for reliability), and easyexpandability. The Most-of the contractors had previous social connections with ARPA program would then systematically vary small portions of this topol­ personnel or contractors. Kleinrock, as has already been noted, ogy, rejecting changes that raised costs or violated constraints and brought . Frank :into the project. In addition, Roberts; Kleinrock, adopting changes that lowered costs without violating constraints. By Heart, and Kahn had all at and Roberts, 60 Chapter 2 61

Table 2.2 .:....~~l.".The IMP messages to The organization of the ARPANET project. .. :.•~bout1000 bits, then add to pae:ket a header cOJltai.nirlg packet's source and destination addresses and some control infor­ ARPNIPTO Project Mamagernent lllation that could be used to check for transmission errors. When the .p;ac-!<.etsreached their destination, the IMP there would strip' off the I .packet headers and reassemble the packets into a complete message Bolt, Beranek University Network Stanford Network . U~.'VJ'·"handing the data over to the host. and Newman of California Analysis Research Working As a packet switch, the IMP had to ensure that data was transmitted at Los Corporation Institute Group »: etlicitmtJly and reliably along each link between a. pair of IMps or Angeles . . bt~tw'eenan IMP and a host. One mechanism for increasing reliability IMP hardware Analysis Topology Netuiark Host was to acknowledge receipt of the packets. Whenever an IMP or a host and software, Information protocols sent a packet across a link, it waited for the recipient to send back a netioork Center operations standard message indicating that the data had been received intact. If this acknowledf,'lnent did not arrive within a given period of time, the sender would transmit the packet again. Before acknowledging receipt Honeywell of a packet, the IMP used a mechanism called a to verify IMP hardware that the data had not been corrupted during Acknowl- edgments and checksums the computing of the IMP to gi¥e links a degree of reliabilitY that many engineers, on the Kleinrock, and most of the members of the BBN team had worked basis of their experience with analog unattain- and become acquainted at the Lincoln Lab. The informal atmosphere able. The IMP was also responsibledo'Fc:onU'QIlirlg of the project was, no doubt, attributable in large paq to the many over the network to prevent BBN team initiaUy social ties among the contractors. H to accomplish this by having the IMP restrict pa(:kets Table 2.2 illustrates the organization8f the i\RPANET project. The any host could send into the netwoi'k at one time; howevel~designing distri~:utionof contracts followed the layered ciivision of the network good flow-control mechanisms proved to be a difficult task that com­ itself. BBN, UCLA, and .tI{ACwork.ed on tpe communications layer, puter scientists were still wrestling with decades later. while the NWG designed the host software and SRI provided docu­ Perhaps the most difficult packet task for the mentation services for the host layer. Having the yarious layers per­ routing. In the ARPANET, muting was <:llist(~bltJJj~Q;:] form independent functions made it easier for ARPA to distribute the a central routing me.chanism, each I~Ed.ecicledl irld~:.pe'ndlenltly development work among several groups. to send packets. 16 To filld tile kept a table with an entry for each host on the network, showing how long it would Managing Technical Complexity take a packet sent from the IMP to re~chthat and which of the IMP's links led to that host by the fastest route. When a packet came Building the IMP: ~uttingLayering into Practice in, the IMP would look up the destination host in the table and At the heart of the communications subnet was the interface message forward the packet via the specified link. The routing system was also processor, which acted both as a packet switch and as an interface notable for being adaptive, continually responding to changes in net­ w. between the host and the network, Since it a..s to be, an interface, , the work configuration or traffic. Every 2/3 second, the IMP would make IMP had to make data from hosts conform to the packet format USed a new estimate of how long it wQuld take to rea,h the various host in the subnet. The task was relatively simple. An IMP would recei"e destinations, and it would send to data from hosts in the form of "messages" that contained up to 8096 neighbors. The IMP used the information sent in its neizhbors 63

on for

the the and

into

net­

each

com­ IMP new

BBN

com­

IMPs,

of the prob­ in the

might

became

request

allowed students

hard

function to see if

An the

software interrne­

changes, the

the hosts

and

automat­

built 121,156).

hardware.

operators.

reload

made

design

be to

distributed

of

by

replace

IMPs, or of

that

to

a

military

technical

would

ARPANET

the

to team

reload

IMP in way.

local

of software

task

of protect

it

operation

IMP

graduate

packets, would to

oflayering

other

neighboring Roberts

to

on

maintenance

it

The network over

unable

of

control at most sites. He was

or the periodically

ideal

in if so,

temperature

the reduce

wOllldrestart

the

that

unreachable

the and program

Lyon 1996, pp. down

technical

independent

suppliers

damaged

relying

so

on to

duplicate control

hosts, on

surges.)

technique

inquisitive

the

IMP was

against

checked

failures

for or

and failure the

from its own failures.

day-to-day proceeded,

the made

or itself

local IMP, especially since com­

the

damaged; the of commodity procedures

form

that

operate

IMP

the lost power

as well as to

such lines,

steps to make

layer shut

eased

their

needforhumanintervention

monitoring power

counting

maintain

been to

"ruggedizingt'

machine recover

(Hafner

a

for

group,

and

a copy

to

Each

behavior

field visits

procedure to

of on his dead

with

The had

self-contained, practical

the

by

than

his diagnostic

solving

destinations

several

robust,

them

relations

remote malfunctionipg

for concerned

subnet

send team

run and

checked and based

IMPs

for the

took

making aresult interfere

to

the

more

automatically

Rather were still a scarce

returned.

common failures. social designed If

program

to as

the

hosts, or communications

destructive a

protecting interference,

Heart

experiment

IMP to

as little as possible on

sort minimized

line

group

IMPs.

the

to by

periodically

particularly without

power

make

allowed the or

network of

down"

it would

radio

any version. members

on manage

tOlamperwith anticipating

on the IMP,

capabilities BBN team gave a

was IMPs always

Designing

operators.

operating

BBN entailed of

was

want

and

depend

IMP

when

staff tested the

IMP also was

ld, work

"went

IMP

basic

The

neighboring

As

~~;i(iOlilst.O ~~ulj,en,ts PI'exlty, .

.~practlce:d ~peration .•..•.• ~uman ~eliably, IMPs

.~",~,q;~'''alll,: l).OU-!unctH:lnl.ng puters, (!iiI{Ulggleditzillg," 'Vibration, the BBN

on an IMP

'fhe ically its a n.ewcopy, (:orrupted work from any network. that cause. By lems, demands

of

the

An

was

rug, net­

had

This The

the

not

deci­ pro­ rout­

basic

BBN

1977, a

robust

layers

should at

subnet

price interac­

with

difficult

network

to revise

complex

the

the

IMP

host.

host com­ simulation

layers.

tbe

subnetwork

would

host more the

computing

had The

commitment

the routing

Walden Hosts would in a

ttle or

Adaptive

how it people

of

described

building

prol':essing

that

that of own local IMP

membel'SJjohn

any

anq

of

team

reliability, the

control of the node

destination

just

particular, netw()rk

for an IMP unexpected

about

was, at times,

essential

belief

host's

goal For

the system

IMP

to message of

high-performance In design

changing designers' must control

reliability

a

would

bYt~am

that

hosts its

experience

autonomous. component.

the

their

effective (McQu~lap

beliefs

or line failures,

and

between each host on

be functions

prone

DlJor

[IMPs] ARPA's

travel time to

unless a

But

to

'resources. network.

Newman's

constantly

node building

a system

functioning strong can be changed under central control anyone

speed reflected

decisions

was

nodes

the

routing made

served

system

depend and

subnet

on

thecQmmunicationsand et al. 1972, p. 244; Stallings 1991). the fundamental

articulated, of and

and the the matching

shortest

switching

contract.

which

distinctions

suhnet

design

creating routing group's tbat

the

and

hierarchy of control, is

routing several times as table,

efficiency,

routes the

of

to

complex ~.

Beranek

packet

ARPA Walden

IMP to

improve independent,

Distributed

(Ornstein

IMP For

than all Host programs.

BBN

to the rest

software,

control.

techniques

dependence

the tl:omanypotentialinterferen(:efrom

the

price

words,

Bolt,

and

maintainability, was a

routing

the the

reported David

its

by

and enforce many

IMPs would not

simply

new

2

functio?s, approach algorithm

congested

of

by For

information,

if its local host went down.

to

approach

relation

other

and

had result, from failures in the

of

separate: isolated

and Hosts must control Host or

weaknesses

a

in

its own on so

interface ....

be be

however,

Chapter

As

tried

BBN vision,

that

specified

disabled.

shaped

routing

avoiding minimizing

exploring

understand

isolated

The impaired

layeri?g the

update

link 62

innovative ged, flexible system.

by ing allowed IMPs to relying by tions. sions,

its revealed

to work, even at the ARPANET's to been

also perform team explanation M(:Quillan

should puters-s-in

should network environment. resources, A subnetwork environment mustprogram be under the gram should be node and much more p.282) In

be were resources be 64 Chapter 2 ARPANET 65 would be used. Though the original plan had been to connect each IMP to a single computer, by 1971 it was apparent that some sites would need to connect multiple computers; thus, ~BN modified the design of the IMP to accommodate multiple hosts (Ornstein et al. 1972, p. 244). In 1971 (two years into the project), Roberts decided to make the network accessible to users whose sites did not have ARPA" By 1970, Roberts was also urging sites to make more use of the NET hosts. He directed BBN to create a new version of the IMp, called year-old network. When sites began turning BBN with questions, a "terminal IMP" or a "TIP," that would interface directly to terminals the Network Control Center (NCG) took on role of providing rather than to hosts. Once connected to a TIP, a user at a terminal network information as well handling trouble reports (McKenzie could access any host on the network. The TIP dramatically extended 1990, p. 20). the community of potential ARPANET users, since it was no longer The responsibility for providing user support fell to Alex McKenzie. necessary for a site to have its own time sharing host. By 1973, half McKenzie had worked at BBN since 1967, but he did not become the sites using the ARPANET were accessing it through TIPs (Roberts involved in the ARPANET project until November 1970, when he 1973, pp. 1-22). returned from an extended vacation looking for a new assignITlent. Frank Heart asked McKenzie if he would be willing to learn enough Operating the Network: Redefining Responsibilities about the IMP to answer questions from the ARPANET sites, so that In September of 1969, representatives from Bolt, Beranek and New­ the IMP team could get on with its development work. McKenzie took man and Leonard Kleinrock's group installed the first IMP at UCL'\. charge of the Network Control Center in 1971, and he responded to This event marked the beginning of the ARPANET's operation, the increasing demands from llsers by expandi q.nd redefining the though the "network" had only one node at this point. By the end of l1g NeG's role ..Believing that BBN should provide ';the reliability of the 1969, less than a year after BBN won its contract, the IMP team power company or the phone company," McKenzie (1990, p. 13), succeeded in installing and linking the four initial nodes at UCL'\, promoted a vision of the ARPANET as a "computer utility." Under his SRI, UC Santa Barbara, and Utah. But although the ARPANET was direction, the NCC acquired a full-time staff and began coordinating able to transmit test messages among the various sites, much work was upgrades of IMP hardware and software..The NCC assumed respon­ still needed before the network could provide a usable communication sibility for fixing all operational problems in the network, whether or system. not BBN's equipn?c:nt was at fault. staff monitored the ARPANET Bolt, Beranek and Newman's contractual responsibilities included Its constantly, recording when each HvlP, line, or host went up or down keeping the IMP subnet running, and Frank Heart's team soon found and taking trouble repprts from users. When monitors detected that operating an experimental distributed network posed its own a disruption of service, they used the IMP's diagnostic features to challenges. BBN set up a Network Control Center in 1970, when the identify its cause. Malfunctions in remote IMPs could often be fixed company's own ARPANET node (the network's fifth) came online. At NC~ first the Network Control Center simply monitored the IMPs, and it from the via the network, using the control functions that BBN had built into the IMPs. was manned "on a rather casual basis" by BBNpersonnel (McKenzie 1976, pp. 6-5). However, as people began using the network, its reli­ The NCC also gave the BBN group additional knowledge of, and ability became an issue, and complaints from users forced BBN to take therefore control over, the telephone network. Heart et al. (1970, P: 565) commented: "From the outset, we viewed the ARPA Network network operations more seriously. IMP and line failures were more common than BBN had anticipated, and, since the effects of a fault in '}S.a systems engineering problem, including portion of the system one location tended to propagate across the network, identifying the supplied by the common . " BBN such expertise diagnosing network troubles NCe able to source of the problem could be difficult. When a network user encoun­ tered trouble, Heart (1990, pp. 5-36) explained, line failures the telephone companies detected them-e-much 66 Chapter 2 ARPANET 67 the carriers' surprise and initial skepticism (Heart 1990, p. 34; Orn­ Dividing the host functions into layers one possibility for stein et al. 1972, p. 253).18 More generally, by building and operating making the task more manageable. AJex McKenzie, who was a member its own switching network ARPA was able to control the characteristics of the Network Working Group, recalled: "We had a concept that of the communications system in such areas as cost, connection setup laverina had to be done, but exactly what the right way to do it was time, error rates, and reliability-areas in which computer users who unclear to us." (McKenzie 1990, p. 8) The NWG's ipitial plan relied on dial-up connections had little say. In this way the ARPANET was to create two protocols: one that would allow users to work inter­ represented a significant step toward integrating computing and tele­ actively on a computer at another site process known as "remote communications systems. login") and one that would transfer files between computers. Both By 1976, the Network Control Center was, according to McKenzie protoccls would occupythe same layer in the network system. Roberts, (1976, pp. 6-5), "the only accessible, responsive, continuously staffed however, noted that both the remote login·protocol and the file trans­ organization in existence which was generally concerned with network fer protocol would have to begin': their operations by setting up a performance as perceived by the user." The initial division between connection between two hosts, and he saw this as a needless duplica­ subnet and host layers had simplified the work of the network's tion of effort. Meeting with the NWG in December of 1969, Lawrence designers; now the NCC allowed the network's users to ignore much Robertstoldthe group to rethink the host protocol.J9Crocker recalled: of the operational complexity of the subnet and to view the entire "Larry made it abundantly dear that our first step was not big enough, communications layer as a black box operated by Bolt, Beranek and and we went back to the drawing hoard." (quoted in Reynolds and Newman. The NCC had become a managerial reinforcement of Postel 1987) ARPA's layering scheme. Roberts suggested separating the hostfunctions into two layers. The first, called the "host layer," wouldfeature a general-purpose protocol Defining Host Protocols: A Collaborative Process to set up communications between;i pair of hosts; the second, called While the team at Bolt, Beranek and Newman was working out the the "applications layer," would specify protocols for network applica­ design and operation of the subnet, the Network Working Group, led tions such as remote login or file transfer (Karp 1973, pp. 270-271). by Stephen Crocker at UCLA, began working; on th~protocol that Having separate host and applications layers would simplify the host would control the host interactions. Members of the NWG were protocol and lessen the burden on the host system's programmers. excited to have the chance to explore fundamental computing issues j,\lso, eliminating the need for each application to duplicate the work of inter-process communications, but they were also daunted by the of setting up a host-to-host connection would make it easier to create lack of prior work in this area, by the complexities of the software applications programs, thereby encouraging people to add to the pool design, and by the need to coordinate the needs and interests of so of network resources. The ARPANET model now had three layers, as many host sites. Having begun with some ambitious ideas, they soon shown in table 2.3. realized that most sites were unwilling to make major changes to their The host-layer protocol, implemented by a piece of software called hosts. They decided that the host protocols would have to be simple. the Network Control Program (NCP), was responsible for setting up Lawrence Roberts, in his earlier work with Thomas Marill, had even connections between hosts. When an application program had data to argued against requiring a network-wide host protocol: "Since the send overthe network, it would calIon the NCp, which would package motivation for the network is to overcome the problems of computer the data into messages and send them to the local IMP. Incoming incompatibility without enforcing standardization, it would not do to messages from the IMP would be collected by the NCP and passed on require adherence to a standard protocol as a prerequisite of mem­ to the designated application. The NCP also made sure that hosts bership in the network." (Marill and Roberts 1966, p. 428) But communicating over the network agreed on data formats. Roberts's earlier experiment had linked only two computers. A net­ The NCP's· design was shaped by assumptions about social and work with dozens of hosts would dearly need some level of stan­ power relations in the networking community. Menlbers NWG dardization to avoid chaos. kept in mind that ARPANET site would have to implement the 68 Chapter 2 ARPAN'ET geller,al use Table 2.3 applications went through contin- The three-layer model of the ARPANET as· NWG and suggested Layer functions One to members used tlle network protc)cols Applications Handles user activities, such as remote login and file transfer the and Host Initiates and maintains connections between pairs of '''''l'vi('FS In the process of working host processes different types of host machines, NWG members Communications Moves data through subnet using packet switching; standing compatibility issues by developing. common. formats for rep­ ensures reliable transmission on host-IMP and resenting files and terminals. These conimon formats became IMP-IMP connections general-purpose tools that aided both networked and non­ networked computers (Crocker et £11.1972, p.

NCP: that is, someone at each site would have to write a program for Managilzg Social Issues the local host computer that would carry out the actions specified by the NCP. Since the host sites were rather reluctant partners in the Lawrence Roberts, who regarded building a sense of community ARPANET development effort, the NCP was designed to be simple, among ARPA's researchers as hoth a means to facilitate network devel- so as to minimize the burden of creating this host software. In addi­ opmentand an end in itself, the ARPANET project tion, members of the Network Working Group were aware that the through a variety of informal aimed at and ARPANET system was being superimposed on existing patterns of reinforcing common values and goals. He maintained personal contact computer use at the various research sites. The NCP's designers were with his contractors through frequent visits, which enabled him to therefore careful to preserve local control over the hosts by Illaking check on the progress of the system and to commitment to remote users subject to the .~amemechanisms for access control, ac­ the project. ARPANET participants could also meet at the. Information counting, and allocation of resources as focal. users (Carr, Crocker, and Processing Techniques Office's annual retreats Jm' Principallnvestiga­ Cerf 1970, p. 591). Finally, th~NWG tried to preserve the auton()~uy tors, praised by Frank Heart (1990, P: 40) as"among the most inter­ of the ARPANET's users,mally of whom were independent-minded esting, .useful meetings that ever took place in the technical computer experts. NWG member Stephen Carr noted: "Restrictions community." Formal presentations..by Pis,.with critiques from their concerning character sets, progranuuing languages, etc., would not be gave'IPTOdirectors an opportunity to the progress of tolerated and we avoided such restrictions." (ibid., p. 79) The NWG their various programs, and the small size of the meetings (generally outlined a design for the NCP early in 1970; by Augllst of 1971, the less than 50 people) was conducive to informal sharing of ideas. protocol had been implemented at all fifteen ARPANET sites. IPTO's assistant director, Barry Wessler, ran similar meetings for the With the host protocol in place, the Network W()rkil1g Group could graduate students working on the ARPANET .(Norberg and O'Neill focus on providing applications. The services originallyenvisioned by 1996, pp. 44-46; Taylor 1989). By bringing researchers from around ARPA, and the first to be put in place, were l:em()~~login and file the United States together to work.en pressingtechnical problems of transfer. Early in 1970, several NWG members devised an experimen­ mutual interest, PI retreats andgraduate student meetings helped the tal remote login program called telnet (for "telecommunications net­ social networks of computer scientists to become national rather than work"), which became generally available in February of 1971 (Croc~er merely local. et £11.1972, p. 273; Carr, Crocker, and Cerf 1970, p. 594). Telnet Good relations among ARPANET contractors initially formed the basis for other services; including file transfer, but nrofessional life more pleasant; were ,n",n,rt,,,",t eventually the NWG created a separate file transfer protocol called ftp. success. In some cases, 70 Chapter 2 Buildiul!: the AR1>AN1~T71 functioning of the network depend on the cooperation of its users. For instance, the first version of the IMP software did not effectively control the flow of data through the network, so it was possible for the system to become overloaded and break down. This was "fixed" in the short term when the users agreed not to send data into the network too fast; their voluntary actions thus compensated for the shortcom­ ings of the system. There were also many points at which it was necessary for the groups at BBN, NAC, and UCLA to coordinate their efforts. BBN, through its monitoring of the network, provided crucial data to the network analysts at NAC and UCLA. There was no other source for data on the performance characteristics of a large distrib­ uted network. Since the theoretical tools of network analysis and simulation were in their infancy, they had to be checked against op­ erational data (Heart et al. 1970, p. 557). Conversely, the analysis and simulation done by NAC and UCLA could aid BBN's engineering wanted to spend-mere time eXpl()Iing- work by predicting potential IMP failures before they appeared in the network. In addition, ~inrpthpcomT)lexityc)f thesvste-rn.rnade resnen­ network. For instance, when Robert Kahn of BBN and the group at sibility for problems difficult UCLA worked together on network simulations, they were able to their own militwilth IDttlers.fiolVaI"d demonstrate that IMPs would be prone to congestion under heavy described NAC as ha'vmgall1 loads." the issue of congestion, This collaborative work was fruitful and often rewarding for those NAC's topology and which Nt\.Cbl;;uI11edon involved, but it also revealed or exacerbated tensions within the com­ Leonard' Kleinrock described tellationsh.ip munity. Despite an ethos of collegiality, there.was alsoa good deal of of "guarded respect," adding that potential conflict among contractors. Bolt, Beranek and Newman was showing up their faults and telling them to at the center of many disputes. The company had much in common pp.25-26). the researchers with ARPA's academic research sites: it was oriented toward research, responsible for analyzing network' performance; and its development groups tended to be small and informal (unlike Kleinrock (ibid.,' p. ,)4.,)rlp~(·ril-,prl· those at many large computer companies). The main IMP team had "competitive but cOoperative." only five members, and the IMP software was designed, programmed, These tensions reflected the fact that the various groups involved in and debugged by three programmers (Heart etal. 1970,p. 566). But the ARPANET project had conflicting priorities fer allocating their BBN was also very much a business, with an eye toward future profits. own time and effort. For instance, when predicted BBN's ARPANET contract represented a chance to get an early start that the performance of the s\lbnet future in the new business of networking (in fact, the company would gain heavy loads, Kleinrock urged· the IMfgroup' to considerable revenues in later years from selling network services). To right away, whereas the BBN team preferrecl. the basic system preserve its strategic advantage in having designed the IMP,BBN installed and functioning·' before,' making improvements. tended to treat the IMP's technical details as trade secrets. In addition, Stephen Crockerrecalled that tIle NWG'sfirst meeting Heart was worried that, if BBN shared too much information about February of 1969 Wi:l.S awkward because of djEferellt'status the IMp, graduate students at the host sites would try to make un­ priorities of the two groups: authorized "improvements" to the IMP software and would wreak I don't thin]; havoc on the system. One of the more heated conflicts within the by Frank ARPANET community arose when BBN refused to share the source 72 Chapter2 A.RR4NET 73 themselves talking to a crewof graduate students they hadn't anticipated. to nr,"""tp svs,telll-',¥ICle we found ourselves talking to people whose first concern was,how to get to flow quickly and reliably but hadn't-e-ofcourse...... spent any time conSciderin~ the thirty or forty layers of protocolabove the link level. (Reynolds and Postel 1987) ltreservillg Illform.al#y: .T~ NetworkW~.vki'llgGrf)Up Within BBN, there was tension between Alex M~Kenzi~'sgroup at the Network Control Center, whose priority wa~tokeep the system up and

w).ders~;aqd running reliably, and the IMP developers, Who wanted to Grol~P·In aS~;lgl:w,Jlg what was behind network malfunctions so as tp pr~ventrecurrences, Roberts had an importaat as,peCt When an IMP failed, the development tean). woulq often keep it ofrelatively inexperienct:d reimal'cJlet;S.\i' rnten of commission for several hours while they d~puggedit, rather than student at UCLA, described it immediately restoring it to service. Heart (1990) commented that th~ and we expecting that some authority wouldfinaHy come along IMP developers came under increasing pres~ureas the n~twq.l'~ and say, 'Here's how we are it! And nobody ever came expanded and became more heavily used: "PeqpI~peg£l.nto qe.pe.p.q along." (Cerf 1990, p. 110) Stephen few upon it. And that was a problem, because that m~

Though ARPA unquestionably played an important role in advanc­ ing basic computer research in the United States, the agency computer science Were proposed careful to present Congress with pragmatic economic or security rea­ themselves, or were designed to allow researchers to continue work in sons for all its projects. It often characterized the ARPANET as an they had explored independently. Of IPTO; (1989, administrative tool for the military rather than as an experiment in pp. 1O~11)said this: computer science. For instance, in 1969 ARPA director Eberhardt constrained to fund something only because of Rechtin promised Congress that the ARPANET "could make a factor relevance.... When I convinced Herzfeld, who was head of 10 to 100 difference in effective computer capacity per dollar among that I Wanted to start the had to take money away the users" (US Congress 1969, p. 809). Two years later; ARPA's new from some other' part of ARPA this thing ·off the ground, he didn't director, Stephen Lukasik, cited "logistics data bases, force levels, and a defense rati1on.aJle. various sorts of personnel files" as military information sources that tven iftlle resultlngtech~ologieseventuallybecame part of the mili­ would benefit from access to the ARPANET (US Congress 1971, tqry comman~and control systeIil,the defense ration~IZmight come p. 6520). Once the ARPANET was up and running, Lukasik (1973,

this way ARPA was able to generate support from both its political and community and an over its research constituencies. 1972).. By the the Communications opened, enough programs were ready to capture the Launching the System attention of the crowds. The thousand or so people who traveled to Washington for the By the end of 1971 most of the infrastructure for the ARPANET was ICCC were able to witnessa remarkable technological feat. From a in place. The fifteen original sites were all connected tq the network, demonstration area containing dozens of computer terminalsratten­ which had begun to expand beyond the ARPA community to include dees were able to use the ARPANET to access computers located sites run by the Air Force and the National Bureau of Standards. But hundreds or thousands of miles away; there was even a temporary link most sites on the network were only minimally involved in resource to Paris. Softwar~onthese computers allow~dparticipants to tryout sharing: the ARPANET had not brought about the radical jump in meteorological models, an a,ir traffic simulatOl; conferencing systems, productivity that had been anticipated. Though the hardware and a mathematics system, experimental databases, a systemfor displaying software developed for the system represented a, great technical Chinese characters, a compueeriaed-ehess player, Joseph Weizen· achievement, the network as a whole could hardly be considered a baurn's psychiatrist program Eliza, and a variety of other applications success if no one used it. (Roberts and K.

. be Its

in

dis­

pm­

and

their

1988, done

.

that major

gradu­ ARPA­

on the

reliable in com­ careful

systems.

Detailed could

a whole

and

computer use,

had of

the

its

et al. 1972, contractors as a based

their

journals reference

experienced at

that

train

local taught

ARPANET all

ensured (Roberts

working

turn

and

findings

Roberts

number that

to

commercial

to

single

would

Ornstein conferences. switching a

understand,

a while the basic computer in forums,

supported

personnel their

at

Investigators,

ARPANET

and

had that 1970;

packet

Lawrence

insistence

the

contractors became

.

committed publish

papers acquired its

communications ARPANET

its

scientists to to

Principal

professional professional others

implemented and ensured

Roberts had

work networks, uses, as were The

around in data of

the

legitimized and

present

techniques.

and

in

for they

to

and project

centers computer

formed

encouraged

students,

contractors

funding

discussed

computer

technology. Kleinrock's

commercial them its

that

Kleinrock

to expertise

for

networking

network ARPANET

for alternative

research

departments, ARPA's

American techniques

on graduate that

ARPAalso

the

would be

of its

of

et ak 1972, p. Leonard

of the theory in

its new

ARPANET 9

community

encouraged

funding

science

experience careers

economic Telenet,

the part

students

computing

149)3

The

ICT'oc][{er

R~ techniques puter vided ARPA accounts seminated and p. queuing ate later NET ARPANET:3(/ with cultivation take US with generation advocate

it

to

of to

by of

the

be;­ the

of

the

start the new

the

cost­ tIle

and for a

started to The able

devel­

to enyi­ quickly

layer­

ARPA­

gwup out momh

afterward builders

reach technical

maintain

growing a

These response justification,

the

invent version

complexity were the to culmination of Communica­ ARPA the use of

ARPANET

reliable,

EEN inc.orporate

provided

ill

in programmers

been 1972 challenges, not rates .. the

of

networks.

.left 19.75.

worked

they companies

highligllting the

ne.tw()rked

conrraerors range

network.subsidiary, had

popularized did

was Q,ad

Creating 1:lY

demonstration,

network point technical kind these

with some and

contractors, partic~l'tr

service

system

percent in a enthusiastic other

with own.

August of

systems to

gr()wth IPTQc.o111111Unity Roberts

the

the

be

67 a wide of ARPA

which

the its own one had,

success

host

telephone

switching, layering. the

face their

by

turning and

among

ARPANET

netw()rks,zUn ARPANET

as high a c<~mputers ARPANET

of

the public

reduce the ARPANET The

should

made

Of

the was the first

own cOlnpany, Packet operating

the which

experts builders US cities in n

In

packet network, experience such some

addition, the presented

and

launching of jumped on the ability of the system:s

and

that ARPANET-like

of dynamic

before marked implications. for of

their ARPANET's general

by this vision a reality. switching clesig

networks

that

and ).In seven

solutions. Telenet network

an Corporation,

work in

debut the redesign

the maintained idea interest.

a

month,

social years

form

task

machinesencountered 5-21

make lasting

model of Congress,

techniques

packet

depended some

encouraged technique

per

traffic on the

few

and defection pp. public

what

intense switching had

service to

market how to

of the a significance

communications to offer the

from

functions, president.

however, would of

of

demonstration 2

existing

to influenced

communications project conflicts

the through Packet

1976,

offering

only percent to this

formidable

left BEN to that packet these

collaborative prominent

began years also that the

commercial

and Communications data be a initiating ICCC

support networking strategies a

Inc., and skeptical

triumphant

layering; Telenet's conference

few. a vision

was a

demonstration

of

The

as a demonstrated

several

The foster

Though been appreciate would planning

ARPANET: only a few These the (Schelonka the the first

engineers tions, responded

Telenet come marker, networks efficient select

of oped techniques NET

obstacles success of to financial of the system

idea ing layering

ARPANET difficulties ronrnent had communications