Adatagram is a basic transfer unit associated with apacket-switched network. Datagrams are typically structured inheader andpayload sections. Datagrams provide aconnectionless communication service across a packet-switched network. The delivery, arrival time, and order of arrival of datagrams need not be guaranteed by the network.

In the early 1970s, the termdatagram was created by combining the wordsdata andtelegram by theCCITT rapporteur on packet switching,^[1]Halvor Bothner-By.^[2][3] While the word was new, the concept had already a long history.

In 1964,Paul Baran described, in aRAND Corporation report, a hypothetical military network having to resist a nuclear attack. Small standardizedmessage blocks, bearing source and destination addresses, werestored and forwarded in computer nodes of a highly redundant meshed computer network. Baran wrote: "The network user who has called up avirtual connection to an end station and has transmitted messages ... might also view the system as a black box providing an apparent circuit connection".^[4] The concept of what we now call avirtual circuit appears in the design,^[5] although no network was built.

In 1967,Donald Davies published a seminal article in which he introduced thepacket andpacket switching. His proposed core network is similar to the one proposed by Paul Baran though developed independently. He assumes that "all users of the network will provide themselves with some kind of error control". His target is a "common-carrier communication network". To support remote access to computer services by user terminals, which at that time were transmitted character by character, he included, at the network periphery, interface computers that convert character flows into packet flows and vice versa.^[6] Davies wrote: "we were really rather against the virtual circuit, because we believed that a communication network should only concern itself with packets, and that any protocols involved in assembling these packets should be done end-to-end, between the customers themselves."^[5]

In 1970, Lawrence Roberts and Barry D. Wessler published an article aboutARPANET, the first multi-node packet-switching network.^[7] An accompanying paper described its switching nodes (theIMPs) and its packet formats.^[8] The network core performed datagram switching as in Baran's and Davies' model, but the service offered to hosts by the network wasconnection oriented.^[9][10] A reliable message transfer service was thus offered to user computers, thus greatly simplifying the network design. This made the ARPANET what would come to be called avirtual circuit network.^[11]

Roberts presented the idea of packet switching to the communication professionals and faced anger and hostility. Before ARPANET was operating, they argued that the router buffers would quickly run out. After the ARPANET was operating, they argued packet switching would never be economic without the government subsidy. Baran faced the same rejection and thus failed to convince the military to construct a packet-switching network.^[12]

In 1973,Louis Pouzin presented his design forCYCLADES, the first large-scale network implementing the pure Davies datagram model.^[13] The CYCLADES team has thus been the first to tackle the highly complex problem of providing user applications a reliablevirtual circuit service^[14] while using theend-to-end principle in a network service known to possibly produce non-negligible datagram losses and reordering.^[15] Although Pouzin's concern "in a first stage is not to make breakthrough [sic] in packet switching technology, but to build a reliable communications tool for Cyclades",^[13] two members of his team,Hubert Zimmerman andGérard Le Lann, made significant contributions to the design of Internet's TCP thatVint Cerf, its main designer, acknowledged.^[16]

In 1981, the Defense Advanced Research Projects Agency (DARPA) issued the first specification theInternet Protocol (IP). It introduced a major evolution of the datagram concept:fragmentation.^[17] With fragmentation, some parts of the global network may use large packet size (typicallylocal area networks to minimize processing overhead), while some others may impose smaller packet sizes (typicallywide area networks to minimize response time). Network nodes may fragment a datagram into several smaller packets.

In 1999, theInternet Engineering Task Force (IETF) sanctioned the use of the already largely deployednetwork address translation (NAT) whereby each public address can be shared by several private devices.^[18] With it, the forthcomingInternet Address exhaustion was delayed, leaving enough time to introduceIPv6, the new generation of Internet Protocol supporting longer addresses. The initial principle of fullend to end network transparency to datagrams was for this relaxed: NAT nodes had to manage per-connection states, making them in partconnection oriented.

In 2015, theIETF upgraded itsinformational 1998 RFC 2309^[19] that datagram switching nodes performactive queue management, to make it a stronger and more detailedbest current practice recommendation through the publication of RFC 7567. While the initial datagram queueing model was simple to implement and needed no more tuning than queue lengths, support of more sophisticated and parametrized mechanisms were found necessary "to improve and preserve Internet performance" (RED,ECN etc.). Further research on the subject was also called for, with a list of identified items.^[20]

The termdatagram is defined as follows:^[21]

"A self-contained, independent entity of data carrying sufficient information to be routed from the source to the destination computer without reliance on earlier exchanges between this source and destination computer and the transporting network."

— RFC 1594

A datagram needs to be self-contained without reliance on earlier exchanges because there is no connection of fixed duration between the two communicating points as there is, for example, in most voice telephone conversations.^[22]

Datagram service is often compared to a mail delivery service; the user only provides the destination address but receives no guarantee of delivery, and no confirmation upon successful receipt. Datagram service is therefore consideredunreliable. Datagram service routes datagrams without first creating a predetermined path. Datagram service is therefore consideredconnectionless. There is also no consideration given to the order in which it and other datagrams are sent or received. In fact, many datagrams in the same group can travel along different paths before reaching the same destination in adifferent order.^[23]

Each datagram has two components, aheader and a datapayload. The header contains all the information sufficient for routing from the originating equipment to the destination without relying on prior exchanges between the equipment and the network. Headers may include source and destination addresses as well astype and length fields. The payload is the data to be transported. This process of nesting data payloads in a tagged header is calledencapsulation.

TheInternet Protocol (IP) defines standards for several types of datagrams. Theinternet layer is a datagram service provided by an IP. For example,UDP is run by a datagram service on the internet layer. IP is an entirely connectionless, best effort, unreliable, message delivery service.TCP is a higher-level protocol running on top of IP that provides a reliable connection-oriented service.

• Datagram socket
• Frame (networking)
• Protocol Wars

• Units of information
• Packets (information technology)

• CS1 French-language sources (fr)
• Articles with short description
• Short description is different from Wikidata