TCP/IP provides end-to-end connectivity specifying how data should be formatted, addressed, transmitted, routed and received at the destination. It has four abstraction layers, each with its own protocols. From lowest to highest, the layers are:
The Internet protocol suite resulted from research and development conducted by the Defense Advanced Research Projects Agency (DARPA) in the early 1970s. After initiating the pioneering ARPANET in 1969, DARPA started work on a number of other data transmission technologies. In 1972, Robert E. Kahn joined the DARPA Information Processing Technology Office, where he worked on both satellite packet networks and ground-based radio packet networks, and recognized the value of being able to communicate across both. In the spring of 1973, Vinton Cerf, the developer of the existing ARPANET Network Control Program (NCP) protocol, joined Kahn to work on open-architecture interconnection models with the goal of designing the next protocol generation for the ARPANET.
By the summer of 1973, Kahn and Cerf had worked out a fundamental reformulation, where the differences between network protocols were hidden by using a common internetwork protocol, and, instead of the network being responsible for reliability, as in the ARPANET, the hosts became responsible. Cerf credits Hubert Zimmerman and Louis Pouzin, designer of the CYCLADES network, with important influences on this design.
The network's design included the recognition it should provide only the functions of efficiently transmitting and routing traffic between end nodes and that all other intelligence should be located at the edge of the network, in the end nodes. Using a simple design, it became possible to connect almost any network to the ARPANET, irrespective of their local characteristics, thereby solving Kahn's initial problem. One popular expression is that TCP/IP, the eventual product of Cerf and Kahn's work, will run over "two tin cans and a string."
A computer, called a router, is provided with an interface to each network. It forwards packets back and forth between them. Originally a router was called gateway, but the term was changed to avoid confusion with other types of gateways.
From 1973 to 1974, Cerf's networking research group at Stanford worked out details of the idea, resulting in the first TCP specification. A significant technical influence was the early networking work at Xerox PARC, which produced the PARC Universal Packet protocol suite, much of which existed around that time.
DARPA then contracted with BBN Technologies, Stanford University, and the University College London to develop operational versions of the protocol on different hardware platforms. Four versions were developed: TCP v1, TCP v2, TCP v3 and IP v3, and TCP/IP v4. The last protocol is still in use today.
In 1975, a two-network TCP/IP communications test was performed between Stanford and University College London (UCL). In November, 1977, a three-network TCP/IP test was conducted between sites in the US, UK, and Norway. Several other TCP/IP prototypes were developed at multiple research centers between 1978 and 1983. The migration of the ARPANET to TCP/IP was officially completed on flag day January 1, 1983, when the new protocols were permanently activated.
In March 1982, the US Department of Defense declared TCP/IP as the standard for all military computer networking. In 1985, the Internet Architecture Board held a three-day workshop on TCP/IP for the computer industry, attended by 250 vendor representatives, promoting the protocol and leading to its increasing commercial use.
In 1985 the first Interop conference was held, focusing on network interoperability via further adoption of TCP/IP. It was founded by Dan Lynch, an early Internet activist. From the beginning, it was attended by large corporations, such as IBM and DEC. Interoperability conferences have been held every year since then. Every year from 1985 through 1993, the number of attendees tripled.Suleras:Citation needed
IBM, ATT and DEC were the first major corporations to adopt TCP/IP, despite having competing internal protocols (SNA, XNS, etc.). In IBM, from 1984, Barry Appelman's group did TCP/IP development. (Appelman later moved to AOL to be the head of all its development efforts.) They navigated the corporate politics to get a stream of TCP/IP products for various IBM systems, including MVS, VM, and OS/2. At the same time, several smaller companies began offering TCP/IP stacks for DOS and MS Windows, such as the company FTP Software, and the Wollongong Group. The first VM/CMS TCP/IP stack came from the University of Wisconsin.
Back then, most of these TCP/IP stacks were written single-handedly by a few talented programmers. For example, John Romkey of FTP Software was the author of the MIT PC/IP package. John Romkey's PC/IP implementation was the first IBM PC TCP/IP stack. Jay Elinsky and Oleg Vishnepolsky of IBM Research wrote TCP/IP stacks for VM/CMS and OS/2, respectively.
The spread of TCP/IP was fueled further in June 1989, when AT&T agreed to put into the public domain the TCP/IP code developed for UNIX. Various vendors, including IBM, included this code in their own TCP/IP stacks. Many companies sold TCP/IP stacks for Windows until Microsoft released its own TCP/IP stack in Windows 95. This event was a little late in the evolution of the Internet, but it cemented TCP/IP's dominance over other protocols, which eventually disappeared. These protocols included IBM's SNA, OSI, Microsoft's native NetBIOS, and Xerox' XNS.Suleras:Citation needed
An early architectural document, RFC 1122, emphasizes architectural principles over layering.
End-to-end principle: This principle has evolved over time. Its original expression put the maintenance of state and overall intelligence at the edges, and assumed the Internet that connected the edges retained no state and concentrated on speed and simplicity. Real-world needs for firewalls, network address translators, web content caches and the like have forced changes in this principle.
Robustness Principle: "In general, an implementation must be conservative in its sending behavior, and liberal in its receiving behavior. That is, it must be careful to send well-formed datagrams, but must accept any datagram that it can interpret (e.g., not object to technical errors where the meaning is still clear)."  "The second part of the principle is almost as important: software on other hosts may contain deficiencies that make it unwise to exploit legal but obscure protocol features." 
The Internet protocol suite uses encapsulation to provide abstraction of protocols and services. Encapsulation is usually aligned with the division of the protocol suite into layers of general functionality. In general, an application (the highest level of the model) uses a set of protocols to send its data down the layers, being further encapsulated at each level.
The "layers" of the protocol suite near the top are logically closer to the user application, while those near the bottom are logically closer to the physical transmission of the data. Viewing layers as providing or consuming a service is a method of abstraction to isolate upper layer protocols from the nitty-gritty detail of transmitting bits over, for example, Ethernet and collision detection, while the lower layers avoid having to know the details of each and every application and its protocol.
Even when the layers are examined, the assorted architectural documents—there is no single architectural model such as ISO 7498, the OSI model—have fewer and less rigidly-defined layers than the OSI model, and thus provide an easier fit for real-world protocols. In point of fact, one frequently referenced document, RFC 1958, does not contain a stack of layers. The lack of emphasis on layering is a strong difference between the IETF and OSI approaches. It only refers to the existence of the "internetworking layer" and generally to "upper layers"; this document was intended as a 1996 "snapshot" of the architecture: "The Internet and its architecture have grown in evolutionary fashion from modest beginnings, rather than from a Grand Plan. While this process of evolution is one of the main reasons for the technology's success, it nevertheless seems useful to record a snapshot of the current principles of the Internet architecture."
RFC 1122, entitled Host Requirements, is structured in paragraphs referring to layers, but the document refers to many other architectural principles not emphasizing layering. It loosely defines a four-layer model, with the layers having names, not numbers, as follows:
Application layer (process-to-process): This is the scope within which applications create user data and communicate this data to other processes or applications on another or the same host. The communications partners are often called peers. This is where the "higher level" protocols such as SMTP, FTP, SSH, HTTP, etc. operate.
Transport layer (host-to-host): The transport layer constitutes the networking regime between two network hosts, either on the local network or on remote networks separated by routers. The transport layer provides a uniform networking interface that hides the actual topology (layout) of the underlying network connections. This is where flow-control, error-correction, and connection protocols exist, such as TCP. This layer deals with opening and maintaining connections between Internet hosts.
Internet layer (internetworking): The internet layer has the task of exchanging datagrams across network boundaries. It is therefore also referred to as the layer that establishes internetworking, indeed, it defines and establishes the Internet. This layer defines the addressing and routing structures used for the TCP/IP protocol suite. The primary protocol in this scope is the Internet Protocol, which defines IP addresses. Its function in routing is to transport datagrams to the next IP router that has the connectivity to a network closer to the final data destination.
Link layer: This layer defines the networking methods within the scope of the local network link on which hosts communicate without intervening routers. This layer describes the protocols used to describe the local network topology and the interfaces needed to affect transmission of Internet layer datagrams to next-neighbor hosts. (cf. the OSI data link layer).
The Internet protocol suite and the layered protocol stack design were in use before the OSI model was established. Since then, the TCP/IP model has been compared with the OSI model in books and classrooms, which often results in confusion because the two models use different assumptions, including about the relative importance of strict layering.
This abstraction also allows upper layers to provide services that the lower layers cannot, or choose not, to provide. Again, the original OSI model was extended to include connectionless services (OSIRM CL). For example, IP is not designed to be reliable and is a best effort delivery protocol. This means that all transport layer implementations must choose whether or not to provide reliability and to what degree. UDP provides data integrity (via a checksum) but does not guarantee delivery; TCP provides both data integrity and delivery guarantee (by retransmitting until the receiver acknowledges the reception of the packet).
This model lacks the formalism of the OSI model and associated documents, but the IETF does not use a formal model and does not consider this a limitation, as in the comment by David D. Clark, "We reject: kings, presidents and voting. We believe in: rough consensus and running code." Criticisms of this model, which have been made with respect to the OSI model, often do not consider ISO's later extensions to that model.
For multiaccess links with their own addressing systems (e.g. Ethernet) an address mapping protocol is needed. Such protocols can be considered to be below IP but above the existing link system. While the IETF does not use the terminology, this is a subnetwork dependent convergence facility according to an extension to the OSI model, the internal organization of the network layer (IONL).
ICMP & IGMP operate on top of IP but do not transport data like UDP or TCP. Again, this functionality exists as layer management extensions to the OSI model, in its Management Framework (OSIRM MF) 
The SSL/TLS library operates above the transport layer (uses TCP) but below application protocols. Again, there was no intention, on the part of the designers of these protocols, to comply with OSI architecture.
The link is treated like a black box here. This is fine for discussing IP (since the whole point of IP is it will run over virtually anything). The IETF explicitly does not intend to discuss transmission systems, which is a less academic but practical alternative to the OSI model.
The following is a description of each layer in the TCP/IP networking model starting from the lowest level.
The link layer is the networking scope of the local network connection to which a host is attached. This regime is called the link in Internet literature. This is the lowest component layer of the Internet protocols, as TCP/IP is designed to be hardware independent. As a result TCP/IP is able to be implemented on top of virtually any hardware networking technology.
The link layer is used to move packets between the Internet layer interfaces of two different hosts on the same link. The processes of transmitting and receiving packets on a given link can be controlled both in the softwaredevice driver for the network card, as well as on firmware or specialized chipsets. These will perform data link functions such as adding a packet header to prepare it for transmission, then actually transmit the frame over a physicalmedium. The TCP/IP model includes specifications of translating the network addressing methods used in the Internet Protocol to data link addressing, such as Media Access Control (MAC), however all other aspects below that level are implicitly assumed to exist in the link layer, but are not explicitly defined.
This is also the layer where packets may be selected to be sent over a virtual private network or other networking tunnel. In this scenario, the link layer data may be considered application data which traverses another instantiation of the IP stack for transmission or reception over another IP connection. Such a connection, or virtual link, may be established with a transport protocol or even an application scope protocol that serves as a tunnel in the link layer of the protocol stack. Thus, the TCP/IP model does not dictate a strict hierarchical encapsulation sequence.
The internet layer has the responsibility of sending packets across potentially multiple networks. Internetworking requires sending data from the source network to the destination network. This process is called routing.
In the Internet protocol suite, the Internet Protocol performs two basic functions:
Host addressing and identification: This is accomplished with a hierarchical addressing system (see IP address).
Packet routing: This is the basic task of sending packets of data (datagrams) from source to destination by sending them to the next network node (router) closer to the final destination.
Some of the protocols carried by IP, such as ICMP (used to transmit diagnostic information about IP transmission) and IGMP (used to manage IP Multicast data) are layered on top of IP but perform internetworking functions. This illustrates the differences in the architecture of the TCP/IP stack of the Internet and the OSI model.
The internet layer only provides an unreliable datagram transmission facility between hosts located on potentially different IP networks by forwarding the transport layer datagrams to an appropriate next-hop router for further relaying to its destination. With this functionality, the internet layer makes possible internetworking, the interworking of different IP networks, and it essentially establishes the Internet. The Internet Protocol is the principal component of the internet layer, and it defines two addressing systems to identify network hosts computers, and to locate them on the network. The original address system of the ARPANET and its successor, the Internet, is Internet Protocol version 4 (IPv4). It uses a 32-bit IP address and is therefore capable of identifying approximately four billion hosts. This limitation was eliminated by the standardization of Internet Protocol version 6 (IPv6) in 1998, and beginning production implementations in approximately 2006.
The transport layer establishes host-to-host connectivity, meaning it handles the details of data transmission that are independent of the structure of user data and the logistics of exchanging information for any particular specific purpose. Its responsibility includes end-to-end message transfer independent of the underlying network, along with error control, segmentation, flow control, congestion control, and application addressing (port numbers). End to end message transmission or connecting applications at the transport layer can be categorized as either connection-oriented, implemented in TCP, or connectionless, implemented in UDP.
The transport layer can be thought of as a transport mechanism, e.g., a vehicle with the responsibility to make sure that its contents (passengers/goods) reach their destination safely and soundly, unless another protocol layer is responsible for safe delivery. The layer simply establishes a basic data channel that an application uses in its task-specific data exchange.
For this purpose the layer establishes the concept of the port, a numbered logical construct allocated specifically for each of the communication channels an application needs. For many types of services, these port numbers have been standardized so that client computers may address specific services of a server computer without the involvement of service announcements or directory services.
For example, the TCP is a connection-oriented protocol that addresses numerous reliability issues to provide a reliable byte stream:
data arrives in-order
data has minimal error (i.e. correctness)
duplicate data is discarded
lost/discarded packets are resent
includes traffic congestion control
The newer Stream Control Transmission Protocol (SCTP) is also a reliable, connection-oriented transport mechanism. It is message-stream-oriented — not byte-stream-oriented like TCP — and provides multiple streams multiplexed over a single connection. It also provides multi-homing support, in which a connection end can be represented by multiple IP addresses (representing multiple physical interfaces), such that if one fails, the connection is not interrupted. It was developed initially for telephony applications (to transport SS7 over IP), but can also be used for other applications.
User Datagram Protocol is a connectionless datagram protocol. Like IP, it is a best effort, "unreliable" protocol. Reliability is addressed through error detection using a weak checksum algorithm. UDP is typically used for applications such as streaming media (audio, video, Voice over IP etc.) where on-time arrival is more important than reliability, or for simple query/response applications like DNS lookups, where the overhead of setting up a reliable connection is disproportionately large. Real-time Transport Protocol (RTP) is a datagram protocol that is designed for real-time data such as streaming audio and video.
The applications at any given network address are distinguished by their TCP or UDP port. By convention certain well known ports are associated with specific applications. (See List of TCP and UDP port numbers.)
Since the IP stack defines no layers between the application and transport layers, the application layer must include any protocols that act like the OSI's presentation and session layer protocols. This is usually done through libraries.
Application layer protocols generally treat the transport layer (and lower) protocols as black boxes which provide a stable network connection across which to communicate, although the applications are usually aware of key qualities of the transport layer connection such as the end point IP addresses and port numbers. As noted above, layers are not necessarily clearly defined in the Internet protocol suite. Application layer protocols are most often associated with client–server applications, and the commoner servers have specific ports assigned to them by the IANA: HTTP has port 80; Telnet has port 23; etc. Clients, on the other hand, tend to use ephemeral ports, i.e. port numbers assigned at random from a range set aside for the purpose.
Transport and lower level layers are largely unconcerned with the specifics of application layer protocols. Routers and switches do not typically "look inside" the encapsulated traffic to see what kind of application protocol it represents, rather they just provide a conduit for it. However, some firewall and bandwidth throttling applications do try to determine what's inside, as with the Resource Reservation Protocol (RSVP). It's also sometimes necessary for Network Address Translation (NAT) facilities to take account of the needs of particular application layer protocols. (NAT allows hosts on private networks to communicate with the outside world via a single visible IP address using port forwarding, and is an almost ubiquitous feature of modern domestic broadband routers).
Layer names and number of layers in the literatureEdit
The following table shows various networking models. The number of layers varies between three and seven.
The three top layers in the OSI model—the application layer, the presentation layer and the session layer—are not distinguished separately in the TCP/IP model where it is just the application layer. While some pure OSI protocol applications, such as X.400, also combined them, there is no requirement that a TCP/IP protocol stack must impose monolithic architecture above the transport layer. For example, the NFS application protocol runs over the eXternal Data Representation (XDR) presentation protocol, which, in turn, runs over a protocol called Remote Procedure Call (RPC). RPC provides reliable record transmission, so it can run safely over the best-effort UDP transport.
Different authors have interpreted the RFCs differently, about whether the link layer (and the TCP/IP model) covers OSI model layer 1 (physical layer) issues, or if a hardware layer is assumed below the link layer.
Several authors have attempted to incorporate the OSI model's layers 1 and 2 into the TCP/IP model, since these are commonly referred to in modern standards (for example, by IEEE and ITU). This often results in a model with five layers, where the link layer or network access layer is split into the OSI model's layers 1 and 2.
The session layer roughly corresponds to the Telnet virtual terminal functionalitySuleras:Citation needed, which is part of text based protocols such as the HTTP and SMTP TCP/IP model application layer protocols. It also corresponds to TCP and UDP port numbering, which is considered as part of the transport layer in the TCP/IP model. Some functions that would have been performed by an OSI presentation layer are realized at the Internet application layer using the MIME standard, which is used in application layer protocols such as HTTP and SMTP.
The IETF protocol development effort is not concerned with strict layering. Some of its protocols may not fit cleanly into the OSI model, although RFCs sometimes refer to it and often use the old OSI layer numbers. The IETF has repeatedly statedSuleras:Citation needed that Internet protocol and architecture development is not intended to be OSI-compliant. RFC 3439, addressing Internet architecture, contains a section entitled: "Layering Considered Harmful".
Conflicts are apparent also in the original OSI model, ISO 7498, when not considering the annexes to this model (e.g., ISO 7498/4 Management Framework), or the ISO 8648 Internal Organization of the Network layer (IONL). When the IONL and Management Framework documents are considered, the ICMP and IGMP are neatly defined as layer management protocols for the network layer. In like manner, the IONL provides a structure for "subnetwork dependent convergence facilities" such as ARP and RARP.
IETF protocols can be encapsulated recursively, as demonstrated by tunneling protocols such as Generic Routing Encapsulation (GRE). GRE uses the same mechanism that OSI uses for tunneling at the network layer.
No specific hardware or software implementation is required by the protocols or the layered model, so there are many. Most computer operating systems in use today, including all consumer-targeted systems, include a TCP/IP implementation.
A minimally acceptable implementation includes the following protocols, listed from most essential to least essential: IP, ARP, ICMP, UDP, TCP and sometimes IGMP. In principle, it is possible to support only one transport protocol, such as UDP, but this is rarely done, because it limits usage of the whole implementation. IPv6, beyond its own version of ARP (NBP), ICMP (ICMPv6) and IGMP (IGMPv6), has some additional required functions, and often is accompanied by an integrated IPSec security layer. Other protocols could be easily added later (possibly being implemented entirely in userspace), such as DNS for resolving domain names to IP addresses, or DHCP for automatically configuring network interfaces.
Normally, application programmers are concerned only with interfaces in the application layer and often also in the transport layer, while the layers below are services provided by the TCP/IP stack in the operating system. Most IP implementations are accessible to programmers through sockets and APIs.
Microcontroller firmware in the network adapter typically handles link issues, supported by driver software in the operational system. Non-programmable analog and digital electronics are normally in charge of the physical components below the link layer, typically using an application-specific integrated circuit (ASIC) chipset for each network interface or other physical standard. High-performance routers are to a large extent based on fast non-programmable digital electronics, carrying out link level switching.