The details of implementation and semantics of the transport layer of theInternet protocol suite,[1], which is the foundation of theInternet, and theOSI model of general networking are different. The protocols in use today in this layer for the Internet all originated in the development of TCP/IP. In the OSI model, the transport layer is often referred to asLayer 4, orL4,[2] while numbered layers are not used in TCP/IP.
The best-known transport protocol of the Internet protocol suite is theTransmission Control Protocol (TCP). It is used for connection-oriented transmissions, whereas the connectionlessUser Datagram Protocol (UDP) is used for simpler messaging transmissions. TCP is the more complex protocol, due to itsstateful design, incorporating reliable transmission and data stream services. Together, TCP and UDP comprise essentially all traffic on the Internet and are the only protocols implemented in every major operating system. Additional transport layer protocols that have been defined and implemented include theDatagram Congestion Control Protocol (DCCP) and theStream Control Transmission Protocol (SCTP).
Transport layer services are conveyed to an application via a programming interface to the transport layer protocols. The services may include the following features:[4]
Same order delivery: The network layer doesn't generally guarantee that packets of data will arrive in the same order that they were sent, but often this is a desirable feature. This is usually done through the use of segment numbering, with the receiver passing them to the application in order. This can causehead-of-line blocking.
Reliability: Packets may be lost during transport due tonetwork congestion and errors. By means of anerror detection code, such as achecksum, the transport protocol may check that the data is not corrupted, and verify correct receipt by sending anACK orNACK message to the sender.Automatic repeat request schemes may be used to retransmit lost or corrupted data.
Flow control: The rate of data transmission between two nodes must sometimes be managed to prevent a fast sender from transmitting more data than can be supported by the receivingdata buffer, causing a buffer overrun. This can also be used to improve efficiency by reducingbuffer underrun.
Congestion avoidance:Congestion control can control traffic entry into a telecommunications network, so as to avoidcongestive collapse by attempting to avoid oversubscription of any of the processing orlink capabilities of the intermediate nodes and networks and taking resource reducing steps, such as reducing the rate of sendingpackets. For example,automatic repeat requests may keep the network in a congested state; this situation can be avoided by adding congestion avoidance to the flow control, includingslow start. This keeps the bandwidth consumption at a low level in the beginning of the transmission, or after packet retransmission.
Multiplexing:Ports can provide multiple endpoints on a single node. For example, the name on a postal address is a kind of multiplexing and distinguishes between different recipients of the same location. Computer applications will each listen for information on their own ports, which enables the use of more than onenetwork service at the same time. It is part of the transport layer in theTCP/IP model, but of thesession layer in the OSI model.
The transport layer is responsible for delivering data to the appropriate application process on the host computers. This involvesstatistical multiplexing of data from different application processes, i.e., forming data segments, and adding source and destination port numbers in the header of each transport layer data segment. Together with the source and destination IP address, the port numbers constitute anetwork socket, i.e., an identification address of the process-to-process communication. In the OSI model, this function is supported by thesession layer.
Some transport layer protocols, for example TCP, but not UDP, supportvirtual circuits, i.e., provideconnection-oriented communication over an underlying packet-orienteddatagram network. A byte stream is delivered while hiding the packet mode communication for the application processes. This involves connection establishment, dividing of the data stream into packets called segments, segment numbering and reordering of out-of-order data.
UDP is a very simple protocol and does not provide virtual circuits, nor reliable communication, delegating these functions to theapplication program. UDP packets are calleddatagrams, rather than segments.
TCP is used for many protocols, includingHTTP web browsing and email transfer. UDP may be used formulticasting andbroadcasting, since retransmissions are not possible to a large number of hosts. UDP typically gives higherthroughput and shorter latency and is therefore often used for real-time multimedia communication where packet loss occasionally can be accepted, for example, IP-TV and IP-telephony, and for online computer games.
Many non-IP-based networks, such asX.25,Frame Relay andATM, implement connection-oriented communication at the network or data link layer rather than the transport layer. In X.25, in telephone network modems and in wireless communication systems, reliable node-to-node communication is implemented at lower protocol layers.
The OSI connection-mode transport layer protocol specification defines five classes of transport protocols:TP0, providing the least error recovery, toTP4, which is designed for less reliable networks.
Due toprotocol ossification, TCP and UDP are the only widely used transport protocols on the Internet.[6] To avoidmiddlebox intolerance, new transport protocols may mimic thewire image of a tolerated protocol, orbe encapsulated in UDP, accepting some overhead (e.g., due to outer checksums made redundant by inner integrity checks).[7]QUIC takes the latter approach, rebuilding reliable stream transport on top of UDP.[8]
^RUDP is not officially standardized. There have been no standard-related developments since 1999.
^Excluding data chunk headers and overhead chunks. Without embedded chunks, an SCTP packet is essentially useless.
^Counted as follows: 12 bytes SCTP header + 16 bytes DATA chunk header or 20 bytes I-DATA chunk header + 16+ bytes SACK chunk. Additional non-data chunks (e.g. AUTH) and/or headers for additional data chunks, which might easily increase the overhead with 50 bytes or more, not counted.
ISO/IEC 8073/ITU-T Recommendation X.224, "Information Technology - Open Systems Interconnection - Protocol for providing the connection-mode transport service", defines five classes of connection-mode transport protocols designated class 0 (TP0) to class 4 (TP4). Class 0 contains no error recovery and was designed for use on network layers that provide error-free connections. Class 4 is closest to TCP, although TCP contains functions, such as the graceful close, which OSI assigns to the session layer. All OSI connection-mode protocol classes provide expedited data and preservation of record boundaries. Detailed characteristics of the classes are shown in the following table:[11]
Service
TP0
TP1
TP2
TP3
TP4
Connection-oriented network
Yes
Yes
Yes
Yes
Yes
Connectionless network
No
No
No
No
Yes
Concatenation and separation
No
Yes
Yes
Yes
Yes
Segmentation and reassembly
Yes
Yes
Yes
Yes
Yes
Error recovery
No
Yes
No
Yes
Yes
Reinitiate connection (if an excessive number ofPDUs are unacknowledged)
No
Yes
No
Yes
No
Multiplexing and demultiplexing over a singlevirtual circuit
No
No
Yes
Yes
Yes
Explicit flow control
No
No
Yes
Yes
Yes
Retransmission on timeout
No
No
No
No
Yes
Reliable Transport Service
No
Yes
No
Yes
Yes
There is also a connectionless transport protocol, specified by ISO/IEC 8602/ITU-T Recommendation X.234.[12]
^"RFC 8200". RetrievedJune 15, 2025.Unlike IPv4, the default behavior when UDP packets are originated by an IPv6 node is that the UDP checksum is not optional.