Transmission Control Protocol

The Transmission Control Protocol (TCP), documented in RFC 793, makes up for IP's deficiencies by providing reliable, stream-oriented connections that hide most of IP's shortcomings. The protocol suite gets its name because most TCP/IP protocols are based on TCP, which is in turn based on IP. TCP and IP are the twin pillars of TCP/IP.

TCP adds a great deal of functionality to the IP service it is layered over:
  • Streams. TCP data is organized as a stream of bytes, much like a file. The datagram nature of the network is concealed. A mechanism (the Urgent Pointer) exists to let out-of-band data be specially flagged.

  • Reliable delivery. Sequence numbers are used to coordinate which data has been transmitted and received. TCP will arrange for retransmission if it determines that data has been lost.

  • Network adaptation. TCP will dynamically learn the delay characteristics of a network and adjust its operation to maximize throughput without overloading the network.

  • Flow control. TCP manages data buffers, and coordinates traffic so its buffers will never overflow. Fast senders will be stopped periodically to keep up with slower receivers.

Full-duplex Operation

No matter what the particular application, TCP almost always operates full-duplex. The algorithms described below operate in both directions, in an almost completely independent manner. It's sometimes useful to think of a TCP session as two independent byte streams, traveling in opposite directions. No TCP mechanism exists to associate data in the forward and reverse byte streams. Only during connection start and close sequences can TCP exhibit asymmetric behavior (i.e. data transfer in the forward direction but not in the reverse, or vice versa).

Sequence Numbers

TCP uses a 32-bit sequence number that counts bytes in the data stream. Each TCP packet contains the starting sequence number of the data in that packet, and the sequence number (called the acknowledgment number) of the last byte received from the remote peer. With this information, a sliding-window protocol is implemented. Forward and reverse sequence numbers are completely independent, and each TCP peer must track both its own sequence numbering and the numbering being used by the remote peer.

TCP uses a number of control flags to manage the connection. Some of these flags pertain to a single packet, such as the URG flag indicating valid data in the Urgent Pointer field, but two flags (SYN and FIN), require reliable delivery as they mark the beginning and end of the data stream. In order to insure reliable delivery of these two flags, they are assigned spots in the sequence number space. Each flag occupies a single byte.

Window Size and Buffering

Each endpoint of a TCP connection will have a buffer for storing data that is transmitted over the network before the application is ready to read the data. This lets network transfers take place while applications are busy with other processing, improving overall performance.

To avoid overflowing the buffer, TCP sets a Window Size field in each packet it transmits. This field contains the amount of data that may be transmitted into the buffer. If this number falls to zero, the remote TCP can send no more data. It must wait until buffer space becomes available and it receives a packet announcing a non-zero window size.

Sometimes, the buffer space is too small. This happens when the network's bandwidth-delay product exceeds the buffer size. The simplest solution is to increase the buffer, but for extreme cases the protocol itself becomes the bottleneck (because it doesn't support a large enough Window Size). Under these conditions, the network is termed an LFN (Long Fat Network - pronounced elephant).

RFC 1072 discusses LFNs.

Round-Trip Time Estimation

When a host transmits a TCP packet to its peer, it must wait a period of time for an acknowledgment. If the reply does not come within the expected period, the packet is assumed to have been lost and the data is retransmitted. The obvious question - How long do we wait? - lacks a simple answer. Over an Ethernet, no more than a few microseconds should be needed for a reply. If the traffic must flow over the wide-area Internet, a second or two might be reasonable during peak utilization times. If we're talking to an instrument package on a satellite hurtling toward Mars, minutes might be required before a reply. There is no one answer to the question - How long?

All modern TCP implementations seek to answer this question by monitoring the normal exchange of data packets and developing an estimate of how long is "too long". This process is called Round-Trip Time (RTT) estimation. RTT estimates are one of the most important performance parameters in a TCP exchange, especially when you consider that on an indefinitely large transfer, all TCP implementations eventually drop packets and retransmit them, no matter how good the quality of the link. If the RTT estimate is too low, packets are retransmitted unnecessarily; if too high, the connection can sit idle while the host waits to timeout.