ArcLibrary
L1入门L2进阶L3生态

TCP (Transmission Control Protocol)

Connection-oriented, reliable, in-order, with built-in flow and congestion control — the default transport for most of the internet.

TCPTransportReliable
核心 · Key Idea

In one line: TCP sets up a logical connection between two hosts and delivers an upper-layer byte stream reliably and in order — retransmits losses, slows down on congestion. It underpins HTTP/SSH/SMTP/databases and most other protocols.

What it is#

TCP provides:

  • Connection-oriented: 3-way handshake before data;
  • Reliable delivery: sequence numbers + ACKs + retransmits;
  • In-order: receiver reorders to send order;
  • Flow control: sliding window, receiver tells sender "how much I can take";
  • Congestion control: detect loss → slow down (Reno / CUBIC / BBR).

The cost is handshake / retransmit overhead — bad for real-time.

Analogy#

打个比方 · Analogy

TCP is a registered letter:

  • Recipient signs (ACK);
  • Lost in transit → courier resends (retransmit);
  • Must arrive in order (sequence numbers);
  • Recipient overloaded → tells you to slow down (sliding window).

Key concepts#

3-way handshake3-way handshake
SYN → SYN+ACK → ACK to establish a connection. See the TCP handshake page.
4-way close4-way close
FIN/ACK both directions to close.
SEQ / ACKSEQ / ACK
Byte counters; receiver tells sender 'next byte expected'.
Sliding windowSliding window
Receiver dynamically advertises 'still got X bytes' for flow control.
MSSMaximum Segment Size
Max payload bytes per segment, often = MTU - 40.
Congestion controlCongestion control
Reno / CUBIC / BBR pace by loss / RTT.

How it works#

The TCP header is at least 20 bytes — sequence, ack, window, checksum, flags (SYN / ACK / FIN / RST / PSH / URG).

Practical notes#

  • ss -ti prints current congestion algorithm, RTT, cwnd per connection.
  • sysctl net.ipv4.tcp_congestion_control changes default. BBR usually outperforms CUBIC on transoceanic links.
  • Lots of TIME_WAIT: high-concurrency short-connection services accumulate them; tune tcp_tw_reuse or use long connections / pooling.
  • Half-open / accept queue full: when a listener is overrun, SYNs get dropped; tune somaxconn and tcp_max_syn_backlog.
  • TCP keepalive: 7200 s default is too long; long-connection services typically use 60–120 s.

Easy confusions#

TCP
Connection-oriented, reliable, ordered, congestion-controlled.
Handshake + retransmit overhead.
UDP
Connectionless, no guarantees, no order, no congestion control.
Zero overhead — fits real-time.

Further reading#