CCNA 200-301 · 1 of 16

How a network is built: hosts, switches, routers and the layers between them

The pieces a network is made of, the two layer models everyone argues about, what each box actually reads before forwarding, the cables and speeds behind the ports, and the vocabulary the rest of the series assumes.

Updated 2026-09-09
On this page

Two laptops on a desk, joined by one cable, are already a network. Everything after that — the switch in the cupboard, the router with the blinking lights, the fibre running under a road, the layer models drawn on every whiteboard — exists to answer one question at larger and larger scale: given these bytes and a name for where they should end up, which wire do I push them down next? This article builds the answer from one cable up, and hands over the words the other fifteen articles use without stopping to explain.

The map

Read this first when short on time. Every branch is a section below.

Figure 1. The whole article on one page. Every branch is a section below; fold what you know, open what you do not.

The smallest network: two hosts and a wire

A network is a set of hosts joined by links, plus enough agreement about naming and forwarding that any host can reach any other. Start with two.

Each host has a network interface: a chip with a socket on the back of the machine. Plug a cable between two of them and you have a link. The interface has two identities, and confusing them is the single most common beginner mistake in networking.

The useful mental model: the MAC address is the name shouted across a room; the IP address is the postal address on an envelope. Shouting works only in the room you are standing in. The envelope works anywhere, but somebody in each room has to shout the last few metres.

One more identity sits above both. A host runs many programs at once, so the arriving bytes need a third label to say which one wants them: a port number. Web servers listen on 80 and 443, SSH on 22, DNS on 53. An IP address plus a port plus a protocol is a socket, and that triple is what a conversation is actually addressed to.

Laptop interface: a4:2b:8c:11:22:33 address: 192.168.1.10 browser: source port 51288 Server interface: 00:1b:44:aa:bb:cc address: 192.168.1.20 nginx: listening on port 80 one link Cat6, 1 Gbps, 3 metres
Figure 2. Three identities on one host. The MAC address names the interface on this cable, the IP address names the host anywhere, and the port names the program.
In the wild
  • Linux shows all three with ip addr for addresses and ss -tlnp for listening ports.
  • Windows prints the same picture from ipconfig /all, which labels the MAC address "Physical Address".
  • Cisco IOS gives you show interfaces for the MAC and speed, and show ip interface brief for the addresses — the command you will type more than any other.
Host
Any device with an IP address that sends or receives its own traffic: laptop, phone, server, printer, camera. Routers and switches are hosts too when you manage them.
Interface
One network attachment point on a host or device. A laptop typically has two (Wi-Fi and Ethernet); a switch has 24 or 48.
MAC address
A 48-bit link-layer address. The first 24 bits are the manufacturer's OUI, so 00:1b:44 tells you who made the card.
Socket
The combination of protocol, IP address and port that one end of a conversation is bound to.

Layers: why nobody designed this as one program

A layer model splits "move these bytes to that machine" into jobs that can be solved independently, so that changing one of them does not force a rewrite of the rest. That is the whole idea, and it is worth holding on to, because both models below are taught as lists to memorise when they are really a design argument.

The argument goes like this. Moving electricity down a copper pair is a completely different problem from finding a path across the planet, which is a completely different problem from recovering a lost chunk of a file. Solve each one separately, define what each hands to the next, and you can swap Wi-Fi in for copper without touching the web browser, or swap HTTP for SSH without touching the router.

The two models

OSI has seven layers and is the vocabulary everyone speaks. TCP/IP has four and is what the software actually implements. They are not rivals; OSI is the ruler, TCP/IP is the thing being measured.

OSITCP/IPJobUnitExamples
7 Application
6 Presentation
5 Session
ApplicationWhat the two programs say to each other, how it is encoded and encryptedDataHTTP, DNS, SSH, TLS
4 TransportTransportWhich program on the host, and whether delivery is guaranteedSegment (TCP)
Datagram (UDP)
TCP, UDP
3 NetworkInternetAddressing and pathfinding across many linksPacketIPv4, IPv6, ICMP, OSPF
2 Data linkLinkDelivery across one link, and access to the shared mediumFrameEthernet, 802.11, PPP
1 PhysicalBits as voltage, light or radioBitsCat6, fibre, RJ45

Two habits make the models useful rather than decorative. First, when something breaks, name the layer before naming the fix — "the link is up but the IP does not answer" is layer 1 and 2 fine, layer 3 suspect, and that alone cuts the search in half. Second, remember that devices are named by the highest layer they read: a layer 2 switch, a layer 3 switch, a layer 7 load balancer.

Encapsulation: the envelopes

Layering shows up on the wire as nesting. Each layer takes what it was handed, prepends its own header, and passes the result down as opaque payload. The receiving host unwraps in the opposite order.

Ethernet frame · dest a4:2b:8c:11:22:33 · plus a trailing checksum IP packet · 192.168.1.10 to 93.184.216.34 · TTL 64 TCP segment · port 51288 to port 80 · sequence 1 HTTP request GET / HTTP/1.1 · Host: example.com
Figure 3. One request as it leaves the interface. Each layer only reads its own header and treats everything inside as payload — which is exactly why a switch can forward a frame without knowing what HTTP is.
In the wild
  • Wireshark renders Figure 3 literally: one collapsible row per layer, outermost first. Clicking a row highlights its bytes in the hex pane.
  • Cloud load balancers are sold by layer — an AWS Network Load Balancer works at layer 4 and forwards without reading the request, an Application Load Balancer works at layer 7 and can route on the URL path.
  • VPNs are encapsulation used on purpose: WireGuard puts a whole IP packet inside a UDP datagram inside another IP packet, so the inner one travels over networks that have never heard of it.
Watch out

"Layer 8" jokes aside, the model has a real limit: TLS does not fit neatly. It sits above TCP and below HTTP, so it is called layer 5, 6 or "layer 4.5" depending on who is talking. Nothing breaks — the model is a description, not a law.

PDU
Protocol data unit: the name for a layer's unit of data. Frame at layer 2, packet at layer 3, segment at layer 4. Exam questions lean on these names.
Encapsulation
Adding a header (and sometimes a trailer) as data moves down the stack. De-encapsulation is stripping them on the way up.
MTU
Maximum transmission unit: the largest payload one link will carry, 1500 bytes on standard Ethernet. Anything larger must be fragmented or rejected.

What each box actually reads

Every network device does the same thing — take a frame in one port, decide, send it out another — and they differ only in how deep they read before deciding. That one sentence explains almost every difference between them.

flowchart LR
  F["Frame arrives"] --> H{"Device?"}
  H -->|Hub| A["Reads nothing<br/>copies bits to<br/>every other port"]
  H -->|Switch| B["Reads destination MAC<br/>sends out the one port<br/>that owns it"]
  H -->|Router| C["Reads destination IP<br/>picks a next hop,<br/>rebuilds the frame"]
  H -->|Firewall| D["Reads IP and ports<br/>permits or drops,<br/>then forwards"]
Figure 4. The same frame, four depths of reading. Everything else about these devices — cost, latency, what they can filter — follows from how far in they look.

Hub: history worth knowing

A hub is an electrical repeater. Bits in on one port, bits out on all others, no memory and no decisions. Every host therefore hears every frame, and if two hosts transmit at once the signals collide and both must back off and retry. All ports form one collision domain. Hubs have not been sold for twenty years, but the term collision domain survives, and half-duplex behaviour still shows up when autonegotiation goes wrong.

Switch: the workhorse

A switch reads the destination MAC address of each frame and forwards it out exactly one port, using a table it builds by watching source addresses. Because each port is its own collision domain and each link is full duplex, two hosts can transmit simultaneously without interfering. A switch does not stop broadcasts: a frame addressed to ff:ff:ff:ff:ff:ff goes out every port, so a switched network is one broadcast domain — until you cut it up with VLANs. Article 5 takes the switch apart properly.

Router: where networks meet

A router reads the destination IP address, looks it up in a routing table, and sends the packet to the next hop — building a brand new layer 2 frame for the outgoing link, with new source and destination MAC addresses. The IP addresses stay the same end to end; the MAC addresses change at every hop. Routers do not forward broadcasts, so every router interface is a boundary between broadcast domains. That single property is why routers exist as much as pathfinding is.

HubSwitchRouter
Reads up toLayer 1Layer 2 (MAC)Layer 3 (IP)
Forwarding tablenoneMAC address table, learnedRouting table, configured or learned
Collision domainsone for all portsone per portone per port
Broadcast domainsoneone per VLANone per interface
Rewrites the framenonoyes, every hop
Unknown destinationn/aflood out all portsdrop, or use the default route
In the wild
  • Layer 3 switches blur the line on purpose: a Cisco Catalyst 9300 switches within a VLAN in hardware and routes between VLANs in the same hardware, which is why campus networks stopped putting a separate router in every wiring closet.
  • Home routers are four devices in one case — a small switch, a wireless access point, a router and a NAT firewall — which is why the word "router" means something vaguer at home than it does in a rack.
  • Cloud networks keep the model but drop the metal: an AWS VPC subnet is a broadcast domain that has no broadcast, and its "router" is a distributed lookup rather than a box.
Collision domain
The set of interfaces whose transmissions can collide with each other. One per switch port on a modern network, so collisions should be zero.
Broadcast domain
The set of interfaces that receive a broadcast sent by any of them. Bounded by routers, and cut up by VLANs.
Flooding
A switch sending a frame out every port except the one it arrived on, because it does not yet know where the destination lives.

Cables, ports and the speeds printed on them

The physical layer is the least glamorous part of networking and the source of a surprising share of real faults, because it is the only layer where a human being with a pair of pliers is in the loop.

Copper: twisted pair

An Ethernet copper cable holds four pairs of wires, each pair twisted along its length. The twist is not decoration: two wires carrying opposite copies of the same signal pick up identical interference, and the receiver subtracts one from the other, cancelling the noise. Tighter twists cancel better, which is the physical difference between the categories.

CategorySpeed it is rated forDistanceWhere you meet it
Cat5e1 Gbps100 mMost existing office cabling
Cat61 Gbps, 10 Gbps to 55 m100 mStandard for new installs
Cat6a10 Gbps100 mRuns to access points and dense floors
Cat825 or 40 Gbps30 mInside a data centre rack only

The 100 metre limit is not a suggestion. It is the distance within which the electrical signal is still distinguishable from noise and within which collision detection could work in the original design. Beyond it, links come up and then behave erratically, which is far more annoying than a link that simply stays down.

Fibre: light instead of electricity

Fibre carries pulses of light down a glass core, so it does not care about electrical noise, cannot conduct a lightning strike between buildings, and goes much further. Two kinds matter:

On a switch, fibre plugs into an SFP slot rather than a fixed socket: a small removable transceiver that turns electrical signals into light. Buying the right SFP for the right fibre and the right distance is its own small art, and mismatches are common enough that show interfaces status is worth running before blaming anything else.

Straight-through, crossover, and why nobody asks any more

Ethernet over copper transmits on one pair and receives on another. A PC and a switch use opposite pairs, so a plain straight-through cable works. Two devices of the same kind — switch to switch, PC to PC — would both transmit on the same pair, so they historically needed a crossover cable with the pairs swapped. Then auto-MDIX arrived: interfaces detect the mismatch and swap internally. On any switch made this century, either cable works, and the topic survives mainly because exams still ask.

Duplex and speed: the mismatch that hides

Two interfaces on a link must agree on speed and duplex. Autonegotiation does this by exchanging capabilities and picking the best both support. Trouble starts when one side is hardcoded and the other is not: the automatic side cannot hear a negotiation, falls back to half duplex, and you get a link that is up, passes small packets, and collapses under load with late collisions and rising error counters. The symptom looks like a bad application; the cause is one line of configuration. Either hardcode both ends or neither.

Watch out

A duplex mismatch never shows up as a down interface. Look for late collisions climbing on the half-duplex side and FCS errors or runts on the other, and for a link whose throughput is a few percent of its rated speed under load but fine when idle.

In the wild
  • Power over Ethernet pushes up to 90 W down the same copper as the data, which is how a modern office runs access points, cameras and desk phones without a power socket near any of them.
  • Direct attach copper cables — a fixed cable with transceivers moulded on both ends — are the cheap way to join switches inside one rack at 10 or 25 Gbps.
  • Submarine cables are single mode fibre taken to its limit: repeaters every 50 to 100 km, and a handful of fibre pairs carrying most of the traffic between continents.
Autonegotiation
The exchange in which two interfaces advertise their supported speeds and duplex modes and settle on the best shared option.
Late collision
A collision detected after the first 64 bytes of a frame have been sent. On a correctly built link this cannot happen, so it is a reliable fingerprint of a duplex mismatch or an over-long cable.
SFP
Small form-factor pluggable: a hot-swappable transceiver module that lets one switch port take copper, multimode fibre or single mode fibre depending on what you plug into it.

Topologies and the words people draw with

A topology is the shape of the links, and the shapes have names because each one trades cost against resilience differently.

Star is what every office actually is: every host has one cable to a central switch. Cheap, easy to reason about, and the switch is a single point of failure. Mesh gives every device a link to every other — resilient, and impossibly expensive past a handful of nodes, so it appears only between core devices as a partial mesh. Ring passes traffic around a loop and survives one break; it lost the office but survives in metro fibre. Bus, one shared cable with every host tapped onto it, is dead and worth knowing only because it is where collision domains came from.

The campus layers

Scale a star up and it becomes a hierarchy with three named tiers, each with one job.

flowchart TD
  C1["Core<br/>fast, dumb, always up"] --- D1["Distribution<br/>routing, policy, ACLs"]
  C1 --- D2["Distribution<br/>second building"]
  D1 --- A1["Access<br/>desks, phones, APs"]
  D1 --- A2["Access<br/>floor 2"]
  D2 --- A3["Access<br/>floor 3"]
Figure 5. The three-tier campus design. Access switches take user ports, distribution switches route between VLANs and enforce policy, and the core does nothing but move traffic between distribution blocks as fast as possible.

In a small site the distribution and core tiers are the same pair of switches, which is called a collapsed core. In a data centre the hierarchy is replaced by spine and leaf: every leaf switch connects to every spine switch and to nothing else, so any server is exactly two hops from any other. That design exists because data centre traffic mostly runs sideways between servers, whereas campus traffic mostly runs up towards the internet.

Scope words

TermScopeWho owns the links
PANA desk — a phone and its earbudsYou
LANA building or floorYou
MANA cityUsually a provider
WANBetween sites, any distanceA provider, leased
SOHOA home or tiny officeYou, and one ISP box

The distinction that matters in practice is ownership, not distance: on a LAN you control both ends of every cable and bandwidth is effectively free, while on a WAN you rent capacity, pay per megabit, and design around the assumption that it will be slower and less reliable than you would like.

In the wild
  • Google and Meta data centres run Clos fabrics — spine and leaf taken to several layers — because that is the only shape whose bandwidth grows by adding switches rather than by buying bigger ones.
  • SD-WAN products from Cisco, Fortinet and others treat several cheap internet links as one WAN, steering traffic per application, which is how branch offices replaced expensive private circuits.
  • Metro Ethernet sells you what looks like a switch port in two different buildings, so a MAN link can behave like a LAN cable that happens to be ten kilometres long.

One small office, end to end

Put every piece together in the smallest network that has all of them: two laptops, a printer, one switch, one router, one ISP link.

  1. The laptop is configured (by DHCP) with address 192.168.1.10, mask 255.255.255.0, gateway 192.168.1.1, and a DNS server. The mask tells it that 192.168.1.0 to 192.168.1.255 is local and everything else is not.
  2. A browser asks for example.com. DNS returns 93.184.216.34. Not local, so the packet must go to the gateway.
  3. The laptop knows the gateway's IP but not its MAC, so it broadcasts an ARP request: "who has 192.168.1.1?" The switch floods it, the router answers, and the laptop caches the reply.
  4. The laptop builds a frame: destination MAC = the router, destination IP = the server. It hands it to the switch, which reads the MAC, finds the router's port in its table, and forwards it out that one port.
  5. The router strips the frame, reads the destination IP, finds no specific match and uses its default route pointing at the ISP. It rewrites the source address to the office's single public address (NAT), builds a new frame for the ISP link, decrements the TTL and sends it.
  6. Somewhere between 10 and 20 routers later, the packet reaches the server. The reply comes back, the router reverses the NAT translation, the switch delivers the frame, and the browser's socket on port 51288 receives the bytes.

Nothing in that walkthrough is specific to this office. Change the numbers and it describes every request you have ever made. The rest of this series takes each step and asks how it works when there are hundreds of hosts, several buildings, dozens of subnets and someone actively trying to break in.

Tip

When a network problem lands on you, walk the same six steps in order and stop at the first one you cannot prove. Link light, then IP address and mask, then gateway reachable by ping, then DNS resolving, then the remote port answering. Almost every fault is at the first step you skipped.

Recap

  • A network is hosts joined by links, plus agreement about addressing and forwarding.
  • Every host carries three identities: a MAC address for this link, an IP address for everywhere, and port numbers for its programs.
  • Layering exists so each job can change independently; OSI's seven layers are the vocabulary, TCP/IP's four are the implementation.
  • Encapsulation is layering made visible: frame wraps packet wraps segment wraps data.
  • Devices differ only in how deep they read — a hub reads nothing, a switch reads MAC, a router reads IP, a firewall reads ports and beyond.
  • A switch gives one collision domain per port but one broadcast domain overall; a router bounds broadcast domains.
  • IP addresses stay the same end to end; MAC addresses are rewritten at every router hop.
  • Copper is twisted pair up to 100 m, categorised by how well the twist cancels noise; fibre is multimode for short runs and single mode for long ones.
  • Auto-MDIX ended the straight-through versus crossover question, but exams still ask it.
  • A duplex mismatch leaves the link up and kills throughput under load; late collisions are its fingerprint.
  • Star is the real-world shape; three-tier campus and spine-leaf are star scaled up for two different traffic patterns.
  • LAN versus WAN is really about who owns the link and therefore what bandwidth costs.

Questions

Say the answer out loud before opening it.

What is the difference between a MAC address and an IP address, and why does a host need both?

A MAC address identifies an interface on one link and never leaves that link; an IP address identifies a host across the whole internetwork and travels end to end.

  • MAC is 48 bits, burned in by the manufacturer, and used by switches to forward within a broadcast domain.
  • IP is assigned by configuration or DHCP, is hierarchical so routers can aggregate it, and is what routing decisions are made on.
  • The pairing is resolved by ARP, which asks "who has this IP?" and caches the MAC that answers.

Every router hop rewrites the source and destination MAC of the frame while leaving the IP addresses untouched, which is the cleanest way to remember which is local and which is global.

Name the seven OSI layers and give the TCP/IP layer each maps to.

Physical, data link, network, transport, session, presentation, application; TCP/IP folds them into link (1 and 2), internet (3), transport (4) and application (5, 6 and 7).

  • Layer 1 moves bits, layer 2 moves frames across one link, layer 3 moves packets across many links, layer 4 identifies the program and decides whether delivery is guaranteed.
  • Layers 5 to 7 are collapsed in practice because no widely used protocol separates them.

TLS is the awkward case, sitting above the transport but below the application, which is why it gets called layer 5, 6 or "4.5" depending on the source.

What is the PDU at each of layers 2, 3 and 4?

Frame at layer 2, packet at layer 3, segment at layer 4 for TCP and datagram for UDP.

  • Below those, layer 1 deals in bits; above them, everything is just "data".
  • The names matter because troubleshooting language depends on them — "frames are being dropped" and "packets are being dropped" point at different devices.

A frame is the only PDU with a trailer as well as a header: Ethernet appends a frame check sequence so the receiver can discard corrupted frames.

How many collision domains and broadcast domains does an 8-port switch create?

Eight collision domains, one per port, and one broadcast domain across all eight — unless VLANs split it further.

  • Each port is a separate, full-duplex link, so simultaneous transmissions on different ports cannot collide.
  • Broadcast frames are flooded out every port in the VLAN, which is why the whole switch is one broadcast domain by default.
  • Configuring three VLANs turns that into three broadcast domains, still eight collision domains.

A hub with the same eight ports would give one collision domain and one broadcast domain, which is exactly why hubs stopped scaling.

A switch receives a frame whose destination MAC is not in its table. What does it do?

It floods the frame out every port in that VLAN except the one it arrived on.

  • This is called unknown unicast flooding, and it is how the network keeps working while the table is still being learned.
  • When the destination replies, the switch learns its MAC from the source field of that reply and stops flooding for it.
  • Entries age out after five minutes by default, so a quiet host gets flooded to again until it speaks.

Constant flooding for an active host is a symptom worth chasing: it usually means asymmetric routing or a MAC table that is being cleared by topology changes.

Why does a router stop broadcasts when a switch does not?

Because a broadcast is a layer 2 concept and a router makes its decisions at layer 3, where the destination address 255.255.255.255 has no next hop to be sent to.

  • A switch forwards a broadcast frame because that is precisely what the destination MAC of all ones instructs it to do.
  • A router therefore bounds the broadcast domain, which is what keeps ARP and other chatter from scaling with the size of the whole network.

The one deliberate exception is a DHCP relay, where the router is configured to catch a specific broadcast and forward it as a unicast to a server on another subnet.

What does encapsulation mean, and what does the receiving host do with it?

Encapsulation is each layer adding its own header to whatever the layer above handed down; the receiver de-encapsulates by stripping headers in the opposite order.

  • The sending host goes application data, then TCP header, then IP header, then Ethernet header and trailer.
  • Each layer treats what it received as opaque payload, which is what lets a switch forward a frame without knowing anything about HTTP.
  • The receiver's layers each read only their own header to decide who to hand the rest to.

Tunnelling protocols exploit the same mechanism deliberately, wrapping a complete IP packet inside another one so it can cross a network that could not otherwise route it.

Why is Ethernet copper cable limited to 100 metres?

Because beyond that distance the signal has attenuated far enough that the receiver can no longer reliably distinguish it from noise, and the original standard also needed the round-trip time short enough for collision detection to work.

  • The limit is 100 m total, conventionally 90 m of solid-core cable in the wall plus 5 m of patch lead at each end.
  • Exceeding it does not necessarily kill the link; it produces intermittent errors, which is harder to diagnose.
  • Fibre removes the constraint, which is why any run between floors or buildings uses it.

Cat8 trades distance for speed in the other direction, supporting 40 Gbps but only to 30 m, which confines it to a single rack.

What is a duplex mismatch and how would you recognise one?

It is one end of a link running full duplex while the other runs half, usually because one side was hardcoded and the other left on autonegotiation.

  • The link stays up, and small amounts of traffic pass fine, so monitoring often shows nothing wrong.
  • Under load throughput collapses to a small fraction of the rated speed.
  • The half-duplex side counts late collisions; the full-duplex side counts FCS errors and runts.

The fix is to make both ends match — either both hardcoded or both on autonegotiation — and modern practice is to leave both automatic unless a device is known to negotiate badly.

When would you still need a crossover cable?

Almost never, because auto-MDIX detects the pairing and swaps internally, but the rule was: crossover between two devices of the same type.

  • Straight-through for PC to switch, router to switch, and anything to a hub.
  • Crossover for switch to switch, PC to PC, router to router, and PC directly to router.
  • Auto-MDIX is on by default on essentially all current switch hardware, so either cable works.

Auto-MDIX depends on autonegotiation, so hardcoding speed and duplex on both ends can quietly disable it and bring the old rule back to life.

What is the difference between single mode and multimode fibre?

Single mode has a very narrow core so light travels one straight path, giving long distance at higher optic cost; multimode has a wider core, light takes several paths, and cheaper optics work over shorter runs.

  • Multimode is typical inside a building or data centre — hundreds of metres.
  • Single mode is used between buildings, to the provider, and across cities — kilometres upwards.
  • The transceiver has to match the fibre type; a single mode SFP on multimode fibre gives an unreliable or dead link.

The physical cause of multimode's distance limit is modal dispersion: paths of different lengths arrive at slightly different times and smear the pulse until bits blur together.

Describe the three-tier campus design and what each tier is for.

Access switches connect end devices, distribution switches aggregate access switches and apply routing and policy, and the core moves traffic between distribution blocks as fast as it can.

  • Access is where port security, VLAN assignment and PoE live.
  • Distribution is the layer 2 to layer 3 boundary, and where ACLs and route summarisation belong.
  • The core is deliberately simple — no policy, no filtering, just forwarding — so it is never the bottleneck or the outage.

Small sites merge the top two into a collapsed core, and data centres replace the whole hierarchy with spine and leaf because their traffic runs between servers rather than towards the internet.

Why is spine and leaf used in data centres instead of a three-tier design?

Because data centre traffic is mostly server to server, and spine-leaf makes every server exactly two hops from every other with equal bandwidth on every path.

  • Every leaf connects to every spine and never to another leaf, so there is no hierarchy to climb.
  • Capacity is added by adding spines, which scales linearly instead of requiring a bigger core.
  • All links are active, using equal-cost multipath rather than spanning tree blocking half of them.

The three-tier design assumes traffic flows north to south towards the internet, which was true for campuses and stopped being true the moment applications were split across many servers.

A user says "the internet is down". Which layer do you check first, and how do you work upward?

Start at layer 1 and stop at the first thing you cannot prove: link light, then address and mask, then gateway, then DNS, then the destination service.

  • Link light and interface counters settle layers 1 and 2 in seconds.
  • ipconfig or ip addr confirms an address was obtained and that the mask is what you expect.
  • Ping the gateway, then a public IP address, then a name — the first that fails tells you whether it is routing, the internet link, or DNS.

Pinging an address before pinging a name is the single fastest way to separate "the network is broken" from "name resolution is broken", and those two have entirely different owners.

Why does a home router blur the definitions in this article?

Because it is four devices in one case: a small switch, a wireless access point, a router, and a NAT firewall.

  • The LAN ports behave as a switch, so devices on them are one broadcast domain.
  • The WAN port is the router boundary, with the ISP on the other side.
  • NAT rewrites addresses so many internal hosts share one public address.

This is why "router" means something much narrower in a rack, where the switching, routing, wireless and firewall functions are separate devices with separate configurations.

What is the difference between a LAN and a WAN in practical terms?

Ownership of the links, which determines what bandwidth costs and therefore how you design.

  • On a LAN you own every cable, so a gigabit is effectively free and you design for convenience.
  • On a WAN you rent capacity from a provider, pay per megabit, and design to minimise how much crosses it.
  • WAN links also have far higher latency and more variable loss, so protocols that chat back and forth behave badly over them.

This is why caching, compression and replication appear at the WAN boundary and almost never inside a LAN — the economics only bite on one side.