The OSI model splits network communication into seven layers, from physical signals up to the application, so that each layer depends only on the one below it and can be replaced without touching the others. Its practical use is locating a fault by testing from the bottom up: no link light means a physical problem, a link with no reply from a neighbouring address means a data-link or network problem, and a working address with a dead service means an application problem. Every layer adds a header, so a full 1,460-byte TCP payload carries 58 bytes of overhead in an Ethernet frame, and a one-byte keystroke carries 63.
Part of Networking
The seven-layer reference model of network communication, used to divide a network into parts that can be built, tested and repaired separately.
Seven Layers Let You Replace One Part Without Rebuilding The Rest
A network has to do several unrelated jobs at once: turn bits into voltages, decide which machine a frame is for, find a route across several networks, make delivery reliable, and speak a language an application understands. The OSI model gives each job its own layer. A layer is allowed to use only the service of the layer below it and to offer a service to the layer above, so swapping cable for radio changes the bottom layer and leaves a web server untouched.
The Network Fundamentals article lists the seven layers in a paragraph. The table below adds where each layer lives in a settlement network and which article covers it.
| Layer | Name | What it does | Where you meet it | Article |
|---|---|---|---|---|
| 7 | Application | The service the user wants | web, mail, file sharing | HTTP |
| 6 | Presentation | Data format and encoding | text encoding, TLS in textbooks | Encryption Basics |
| 5 | Session | Keeping a conversation open | an open TCP connection, login sessions | TCP Protocol |
| 4 | Transport | Ports and reliable delivery | TCP, UDP | UDP Protocol |
| 3 | Network | Addresses and routing | IP, routers | IP Addressing |
| 2 | Data link | Frames between neighbours | Ethernet, switches | Data Link Layer |
| 1 | Physical | Signals on the medium | cable, fibre, radio | Physical Layer |
The Four-Layer Stack Is What Actually Runs
The model that real software follows is the four-layer TCP/IP stack: link, internet, transport and application. Network Fundamentals gives the mapping: the OSI physical and data-link layers are the TCP/IP link layer, and the OSI session, presentation and application layers are all the TCP/IP application layer. In a real system there is no separate piece of software called the session layer or the presentation layer, which is why layers 5 and 6 are the ones people struggle to point at. Use the seven-layer names to talk about where a fault is, and expect the software to be built in four.
A Fault Low In The Stack Shows Up Everywhere Above It, So Test From The Bottom
A fault at a low layer produces symptoms at every layer above it, so start at the bottom and stop at the first test that fails.
Step 1: Physical (layer 1) Look at the link light on both ends. Ethernet Basics states the rule: no link LED means no physical connection, so check the cable, the connectors and the power of both devices before anything else.
Step 2: Data link (layer 2) If the light is on but nothing answers, check the switch port statistics for errors, as Star Topology describes, and confirm both ends agree on speed and duplex.
Step 3: Network (layer 3) Ping a neighbour by its address, then a machine on another subnet. A reply from the first and not the second points at the default gateway or a route, which Routing Basics covers.
Step 4: Transport and application (layers 4 and 7) If addresses work but a name does not, the fault is name lookup. If a name resolves and the service still fails, test the port. Network Fundamentals puts it as: link up but no IP connectivity is the network layer, and IP working while one application fails is the application layer.
Worked Example: Every Layer Adds Bytes
The frame sizes come from the sibling articles: Packet Structure gives a minimum IPv4 header of 20 bytes, Frame Format gives a 14-byte Ethernet header, a 4-byte check sequence, a 1,500-byte maximum payload and a 46-byte minimum payload. The TCP header is taken as its 20-byte minimum, a figure not stated in the other articles.
Full-size segment. Application data 1,460 bytes. Add the 20-byte TCP header: 1,460 + 20 = 1,480. Add the 20-byte IP header: 1,480 + 20 = 1,500, exactly the Ethernet payload limit. Add the Ethernet header and check sequence: 1,500 + 14 + 4 = 1,518 bytes. Overhead is 20 + 20 + 14 + 4 = 58 bytes. Efficiency is 1,460 / 1,518 = 0.962, or 96.2% of the frame is user data.
One keystroke. Application data 1 byte. With TCP and IP headers: 1 + 20 + 20 = 41 bytes. That is below the 46-byte minimum payload, so the frame is padded by 46 - 41 = 5 bytes, giving 14 + 46 + 4 = 64 bytes on the wire. Overhead is 63 bytes. Efficiency is 1 / 64 = 0.016, or 1.6%.
The lesson is about design, not tidiness: many tiny messages waste most of the link, while bulk transfers are nearly free of overhead. The preamble and the gap between frames come on top of both figures.
Common Mistakes
- Starting the diagnosis at the application — a bad crimp produces an application timeout; testing the web server first wastes the afternoon
- Looking for layers 5 and 6 in the software — they are conceptual; a real stack folds them into the application
- Treating the model as a rule that real protocols obey — some protocols span layers, and the model is a way to reason, not a specification to satisfy
What’s Next
- Learn next: Physical Layer — layer 1 in detail, the part that fails first
- Combine with: Core Protocols — the layer 3 to 7 protocols a settlement actually runs