Requirements for the

Vinton G. Cerf Senior Vice President, Architecture and Technology MCI Communications Corporation Ashburn, VA 20147 U.S.A. [email protected]

The requirements for the Internet emerged in part from g Providing for host-to-host ªpipeliningº so the practical experiences drawn from the development, that multiple packets could be enroute from deployment and use of the ARPANET. was source to destination at the discretion of the one of the principal design architects of the ARPANET participating hosts, if the intermediate net- packet switch (Interface Message Processor) while he works allowed it. worked at Bolt Beranek and Newman. Bob joined the g Gateway functions to allow it to forward Defense Advanced Research Projects Agency in late 1972 packets appropriately. This included inter- and outlined his requirements for network evolution that preting IP headers for routing, handling inter- year [3] (quoted with minor punctuation correction): faces, breaking packets into smaller pieces if necessary, etc. Four ground rules were critical to Kahn's early g The need for end-end checksums, reassembly thinking: of packets from fragments and detection of g Each distinct network would have to stand on duplicates, if any. its own and no internal changes could be re- g The need for global addressing quired to any such network to connect it to g Techniques for host-to-host ¯ow control. the Internet. g Interfacing with the various operating sys- g Communications would be on a best effort tems basis. If a packet didn't make it to the ®nal g There were also other concerns, such as destination, it would shortly be retransmitted implementation ef®ciency, internetwork per- from the source. formance, but these were secondary con- g Black boxes would be used to connect the siderations at ®rst. networks; these would later be called gate- ways and routers. There would be no infor- In our discussions that started in early 1973 on the mation retained by the gateways about the question of networking, Bob and I considered speci®cally individual ¯ows of packets passing through the problem of interconnecting a mobile net- them, thereby keeping them simple and work (PRNET), a multi-access packet satellite network avoiding complicated adaptation and recov- (SATNET) and the wireline ARPANET. Bob's Open Ar- ery from various failure modes. chitecture ideas clearly took root in the design of the TCP g There would be no global control at the oper- (later TCP/IP) protocol [1, 2]. ations level. The requirement that the networks comprising the Other key issues that needed to be addressed Internet NOT be modi®ed led directly to the need for were: gateways (now called routers) and to the need for a global address space orthogonal to any intra-network addressing g Algorithms to prevent lost packets from per- structure. The need for end-to-end reliability led to the manently disabling communications and en- retransmission mechanisms of TCP. Flow control was a abling them to be successfully retransmitted known requirement because not all hosts would be equal from the source. in capacity to send or receive data, and the network could

1 potentially be congested. A great deal of attention was The advent of voice over IP has created new require- paid to the problem of long-delayed packets that might ar- ments for interlinking of the public switched telephone rive at a most inconvenient time for the TCP protocol. network and/or private branch networks with the public Implementation experience with the December 1974 TCP Internet and with private IP networks. The mechanisms speci®cation [1] led directly to the addition of a ªthree- proposed to support this requirement are found in the way handshakeº to establish the initial sequence numbers ENUM extensions to the DNS. In effect, ENUM maps for each direction of the bisymmetric TCP ¯ow. international E.164 telephone numbers into Universal It was understood, if dimly, that classes of service Resource Identi®ers through a potentially iterative system would need support, hence the TCP/IP packet header con- of regular expression evaluations embedded in the DNS. tained ¯ags for ªreliability,º ªspeed,º or both. With re- A new DNS record type, the Naming Authority Pointer gard to speed (timeliness of delivery), the motivation for (NAPTR) is the active component of this design. breaking out the IP protocol from TCP was precisely the The creation of the World Wide Web has underscored carriage of real-time speech over the Internet. This appli- the value of streaming audio and video and real-time pro- cation favored rapid transport over absolute reliability. tocols for interactive services including multi-party role Eventually UDP and higher level real-time protocols playing games. emerged that avoided the use of TCP for streaming audio, The notion of virtualizing computing and storage re- video and real-time gaming. sources on the Internet has led to the concept of the global As the number of networks comprising the Internet GRID and the so-called GLOBUS protocols. Among increased. it became clear that the original 256 network these are SOAP, XML, UDDI and others. design was inadequate so the 32-bit IP address structure A major missing requirement from the initial Internet was altered to allow for a much larger number of net- design was security and authenticity. Although much works (about 2 million) through the introduction of Class effort went into making the system robust against random A, B, C and D networks. Eventually even this method failures, it was not proof against deliberate attacks. There proved inadequate and Classless Inter-Domain Routing was a version designed for use by the military that made (CIDR) was introduced to create additional ¯exibility in heavy use of end-to-end packet encryption and that ver- subnetwork sizing. sion was intended to provide considerable security. How- Not long after the roll out of TCP/IP, it became clear ever the details were and are classi®ed and the public sys- that the routing table (the so-called Host.Txt ®le) would tem does not have these features, yet. not scale to the tens to hundreds of thousands of hosts on the Internet so the table was replaced by a distributed References design for host naming called Domain Name System (DNS) and was developed beginning about 1984 by Paul [1] Cerf, V.G. and Kahn, R.E., ªA Protocol for Packet Network Mockapetris and Jon Postel. This hierarchical system has Interconnectionº, IEEE Transactions on Communication scaled to tens of milliions of hosts in the network. The Technology COM-22(5), pp. 627±641 (May 1974). commercialization of the Internet brought new require- [2] Cerf, V.G., Dalal, Y., and Sunshine, C., ªSpeci®cation of ments to the DNS, particularly with regard to competi- Internet Transmission Control Programº, Request for Com- ments, RFC675, Network Working Group, (December tion, intellectual property protection and dispute resolu- 1974), http://www.faqs.org/rfcs/rfc675.html. tion. [3] Leiner, B.M., Cerf, V.G., Clark, D.D., Kahn, R.E., Klein- The growing Internet created its own requirements for rock, L., Lynch, D.C., Postel, J., Roberts, L.G., and , S. scalable routing and that led to the development and evo- Wolff, ªA Brief History of the Internetº, Internet Society lution of the so-called Border Gateway Protocol. This sys- (ISOC), Reston, VA (2000), tem was designed to be used in conjunction with some http://www.isoc.org/internet/history/brief.shtml#Introduction. care in IP address assignment so as to allow consolidation of the global routing tables to keep their absolute size and update requirements under control. Intranet routing proto- cols such as IS-IS and OSPF re¯ected similar needs for scalability and hierarchical structure. The evolution of email brought a new set of require- ments, some of which were re¯ected in extensions to the DNSÐin particular, the creation of the MX record to allow a single MTA to serve multiple apparent email do- mains.

2