What a Network Is and What It Buys You

Networking Basics · 20 min

Core CS · Computer Networks

A wire is not a network until both ends agree

A network is two or more independent machines joined by a medium so they can exchange data. Four of the five things that takes are objects you can point at. The fifth is the one candidates leave out, and without it the other four move bits that mean nothing.

Send two messages on one link, three ways
Bits arriving is not communication. Communication is bits arriving in a format the far end can read, carrying meanings both ends agreed on beforehand, at a rate the far end can absorb.

01 The idea

Two autonomous machines, and a medium between them

A computer network is two or more autonomous devices connected by a transmission medium so that they can exchange data. Every word in that sentence is load-bearing, and the word that does the most work is autonomous. It means neither device is under the other’s control: each has its own processor, its own memory and its own clock, and either one can be switched off without stopping the other.

That word is what keeps the definition from swallowing everything. A processor and the graphics card in the same machine are connected by a bus and exchange data constantly, and they are not a network, because the graphics card does not run on its own. Two laptops joined by a single Ethernet cable are a network, and there is nothing else in the room. Size is not part of the definition and never has been.

The transmission medium is the physical path the signal travels: twisted-pair copper, coaxial cable, optical fibre, or open air. Note that it carries a signal and not data. Voltage changes, pulses of light, radio waves. Nothing on the medium knows it is carrying the letter H, which is the whole reason the last component exists.

Any data communication needs five components, and this is a list you will be asked to recite: sender, receiver, message, transmission medium, protocol. Four of them are objects. You can photograph a laptop, a server, a cable, and print out the message. The fifth is an agreement held identically in two places, with no physical location at all, and that is precisely why it is the one people forget.

A network needs five things. Four you can point at — the sender, the receiver, the message and the medium between them. The fifth is the protocol, and it is the only one of the five that decides whether the bits that arrive mean anything.
AutonomousNeither device controls the other. Each has its own processor, memory and clock, and either can be powered off without stopping the other. This is what separates a network from a computer’s internal bus.
Transmission mediumThe physical path a signal travels. Guided media confine it inside a cable — twisted pair, coaxial, optical fibre. Unguided media do not — radio, microwave, infrared. Both carry a signal, never letters.
ProtocolAn agreed set of rules with three parts: syntax, the format and order of the bits; semantics, what each field means and what to do about it; timing, when to send and how fast. Nothing on the wire enforces it. Both ends load it before the first bit.

02 Worked example

Five characters, one cable, and the rule that makes them readable

Asha’s laptop sends the five characters HELLO to the college server over a Cat 6 twisted-pair cable. One sender, one receiver, one message, one medium. Each character travels as one 8-bit byte (the ASCII code itself needs only 7 bits, and the eighth is a leading zero here), so exactly 40 bits cross that cable. Four of the five components below are boxes you could walk over and photograph. The fifth is highlighted, and it is drawn last only because it has nowhere on the wire to sit — it stretches across the whole row.

SenderAsha’s laptop. Produces the message and drives the signal onto the medium.
MessageHELLO. Five characters, 8 bits each, 40 bits on the wire.
MediumCat 6 twisted pair. Carries voltage changes, not letters.
ReceiverThe college server. Reads the signal and rebuilds the bits.
ProtocolThe agreement both ends loaded before the first bit. Nothing on the cable holds a copy of it.

The first four nodes are inventory. The fifth is where the exam lives, so here it is broken into its three parts, each one failing on its own while the other four components stay exactly as they are. Same laptop, same server, same cable, same 40 bits, every time.

1
Syntax — the format and order of the bitsBoth ends must agree that a character is 8 bits of ASCII and that the most significant bit goes first. The laptop sends H as 0100 1000. If the server has been told to reassemble each byte least significant bit first, it reads 0001 0010, which is byte 18, a control character, not a letter. Every bit arrived. Not one of them is wrong. The message is unreadable.H = 72 = 0100 1000  ·  read in reverse = 0001 0010 = 18  ·  identical bits, no longer a letter
2
Semantics — what each field means, and what to do about itThe laptop puts a length byte in front: 05, meaning “five characters follow”. If the server’s rulebook says the first byte is a message type, it reads type 5 and then treats H, which is 0x48, as a length of 72. It waits for 72 characters that do not exist. Same bits, same order, opposite meaning.05 48 45 4C 4C 4F  →  laptop: length 5, then HELLO  ·  server: type 5, then a 72-byte message
3
Timing — when to send, and how fastThe laptop’s link runs at 1 Gbps and the server’s at 100 Mbps. 1 Gbps is 1,000 Mbps, so the laptop offers ten bits for every one the server can take away. The server’s buffer fills, and once it is full nine of every ten bits have nowhere to go. Nothing was malformed and nothing was misunderstood. The data is lost anyway.1,000 Mbps in  ·  100 Mbps out  →  10:1, so 9 of every 10 bits are dropped once the buffer is full
All three agreed, and only thenWith syntax, semantics and timing agreed, the server rebuilds HELLO from the same 40 bits and knows the message is complete. Change one line of the rulebook at one end and nothing physical changes anywhere. The message stops meaning what it meant.sender + receiver + message + medium + protocol = communication  ·  remove any one and it is not

Notice what stayed constant through all three failures. The cable was fine. The bits were fine. The only thing that moved was the agreement, and it broke the same message three different ways. That is why the protocol is counted as a component and not as a detail: it is the only one of the five you cannot inspect by opening a cabinet, and the only one that has to be held identically in two places at once.

Keep this example in your head for the rest of the lesson. In section 03 those 40 bits get a delay and a throughput number attached to them, and in the console the same laptop and the same server try to talk in both directions on one link.

03 Mechanics

Four criteria, and what each one is measured in

A network is judged on four things, and an interviewer will expect a unit for the first two. This is where vague answers get pushed on, because performance and reliability are numbers and candidates offer adjectives.

CriterionThe question it answersWhat you actually quoteWhat it trades against
Performance How much data, and how fast? throughput in bits per second, and delay in milliseconds Cost, and each other. Pushing more traffic through a link raises throughput and raises queueing delay at the same time.
Reliability How often does it fail, how long until it is back, and does it survive a disaster? frequency of failure, recovery time after a failure, and robustness in a catastrophe Cost. Redundancy means paying every month for a second path you hope never carries anything.
Security Who may read it, who may change it, and what happens after a breach? protection from unauthorised access and from damage in transit, plus a recovery policy Performance. Encryption costs processing, and a handshake costs round trips before any data moves.
Cost What does it take to build, and what does it take to keep running? hardware, cabling or spectrum, and the administration that never stops All three of the above. This is the criterion that settles arguments.

One honest note on the count, because textbooks differ and interviewers notice. Forouzan lists three criteria — performance, reliability, security. Cost is the fourth that every real deployment adds, and it is the one that decides the design. Name all four and say which book you are quoting if you are asked.

Throughput and delay are different questions, and the difference is easiest to see on Asha’s 40 bits. On a 100 Mbps link, pushing 40 bits onto the wire takes 40 ÷ 100,000,000 seconds, which is 0.4 microseconds. If the server is 300 km away over fibre where the signal travels about 200,000 km per second, the bits then spend 300 ÷ 200,000 seconds in flight, which is 1.5 milliseconds, or 1,500 microseconds. The message spends 0.4 microseconds being transmitted and 1,500 microseconds travelling — 3,750 times longer in the second stage than the first. For a small message over a long link, the number on the invoice is not the number that decides how it feels.

Those four criteria are the bill. Here is what the money buys, which is the other half of the answer and the half candidates skip. Each row is a use that justifies building the thing, and each one arrives with its own cost attached, because a use with no cost is a slogan.

What the network buysThe concrete instanceWhat it costs you
Resource sharingOne printer on the corridor serves forty machines. One licence server, one storage array, one GPU box the whole lab queues for.a shared resource is a queue, and a single point of failure until you buy a second
CommunicationMail, chat, voice and video between people who are not in the same building, and between programs that are not on the same machine.the traffic is now somebody’s to carry, and delay becomes something users can feel
Centralised administration and backupOne directory holds every login. One nightly job backs up everyone’s work, because the work lives on the server rather than on forty desks.whoever holds the centre holds everything, which is where security stops being optional
Reliability through redundancyTwo links to the internet on separate routes, two servers holding the same data. One fails and the other carries the load.you pay twice for capacity you use once
ScalabilityAdding the forty-first machine is a cable and a switch port, not a redesign of anything.ports, addresses and broadcast traffic all grow with it, and eventually the design does have to change
Every one of those five is bought with the fourth criterion. Sharing one printer is cheaper than forty; running two links is more expensive than one. A network design is never “the best network”. It is the cheapest one that clears the performance, reliability and security numbers somebody wrote down before the quotes came in.

05 Cheat sheet

The three modes, and the lists you will be asked to recite

Give the mode, then the direction rule, then a real example. A definition with no example is the answer that gets a follow-up question.

ModeDirections the link carriesReal exampleWhat it costs you
Simplexone, permanentlyA keyboard to its computer. Broadcast television to an aerial.half the conversation is impossible; every reply needs a second, separate link
Half duplexboth, one at a timeA walkie-talkie. Classic Ethernet on a hub, using CSMA/CD.a turnaround delay on every reversal, carrying no data
Full duplexboth, at the same timeA telephone call. A switched Ethernet link with a separate path per direction.two channels; split one channel to make them and each direction gets half the capacity
Five components, and the one that gets forgottenSender, receiver, message, transmission medium, protocol. Candidates list four and stop. The protocol is the one that turns arriving bits into a message, and it has three parts you should name without being asked: syntax, semantics, timing.
Four criteria, with units attachedPerformance — throughput in bits per second, delay in milliseconds. Reliability — how often it fails, how long recovery takes, and how it holds up in a catastrophe. Security. Cost. The first two are numbers; quote them as numbers and the answer stops sounding memorised.
Who writes the rulesTwo families you will keep meeting. IEEE standardises what happens on the wire and in the air: 802.3 is Ethernet, 802.11 is Wi-Fi. The IETF standardises what runs on top and publishes each one as an RFC: IPv4 is RFC 791, TCP was RFC 793 and is now RFC 9293.

06 Where & why

Where the modes and the criteria show up by name

None of this is a diagram that only exists in a textbook. Each item below is a standard with a number, or a product somebody is running right now, and an interviewer can hear the difference between a candidate who has met them and one who has read a list.

Wi-Fi 6 · IEEE 802.11ax
Radio is half duplex, and the standard is mostly about turns

A station cannot transmit and receive on the same channel at the same instant — its own transmitter would drown whatever was arriving. So 802.11 spends most of its complexity deciding whose turn it is, using CSMA/CA. When someone says the Wi-Fi feels slower than the cable, a shared half-duplex channel is a large part of the reason.

Switched Ethernet · IEEE 802.3
Full duplex is why a 1 Gbps port is 1 Gbps each way

A switch works at layer 2 and gives every port its own collision domain, so a full-duplex link has no collisions at all and CSMA/CD is switched off. It does not split the broadcast domain — a broadcast still reaches every port. Separating broadcast domains is a router’s job, at layer 3. A hub, at layer 1, separates neither: every port sits in one collision domain and one broadcast domain.

OpenSSH
Security is a criterion, not a feature bolted on later

Telnet carried the login password across the network as readable ASCII on port 23, and every device on the path could read it. SSH does the same job under encryption on port 22. Nothing about the cable changed when sites migrated. The protocol changed, which is the point: the guarantee lives in the agreement, never in the medium.

Cloudflare
Reliability is measured, not asserted, and redundancy is what it costs

A network that survives a failed data centre does so by having more than one, and by rerouting to the survivor in seconds rather than hours. That is the reliability criterion in all three of its parts: how often it fails, how long recovery takes, and whether it stays up when a whole site goes down. The second data centre is the cost criterion, paid every month for capacity that mostly sits idle.

All four are the same trade wearing different clothes. You buy performance, reliability and security with cost, and the design that wins is the one that clears the numbers somebody wrote down, not the one with the longest feature list. Asked what makes a good network, give the four criteria and a unit for the first two, and you have already said more than the answer that question usually gets.

07 Interview questions

What they actually ask

This is the opening question of almost every networking interview, and it is usually answered with a definition that could have come off a slide. The candidates who get past it can name the five components, give a unit for two of the criteria, and produce one concrete failure on demand.

What is a computer network, in one line?
Two or more autonomous devices connected by a transmission medium so that they can exchange data. Autonomous is the word that carries the definition: neither device controls the other, and either can be switched off without stopping the other. Nothing in the definition mentions size, so two laptops on one cable qualify.
Your definition said “autonomous”. Why does that word matter?
Without it the definition swallows every component inside a single machine. A processor and its graphics card are connected and exchange data continuously, and they are not a network, because the graphics card has no independent existence. The test to say out loud is whether either device keeps running when the other is powered off.
Name the five components of a data communication system.
Sender, receiver, message, transmission medium, protocol. Four of them are objects you can point at. The protocol is an agreement held identically at both ends with no physical location, which is exactly why it is the one candidates leave off the list.
What is a protocol, precisely?
An agreed set of rules governing data communication, in three parts. Syntax is the format and order of the bits. Semantics is what each field means and what the receiver should do about it. Timing is when to send and how fast. Nothing on the medium enforces any of it; both ends load the same rulebook before the first bit moves.
Every bit arrived intact and the message was still useless. How?
The two ends disagreed on the rulebook. If the sender transmits each byte most significant bit first and the receiver reassembles least significant bit first, the letter H, which is 0100 1000, is read as 0001 0010 — byte 18, a control character. Identical bits, identical order on the wire, and no letter at the far end. That is a syntax failure, and semantics and timing fail the same way for their own reasons.
What criteria is a network judged on?
Performance, reliability, security and cost. Performance is throughput in bits per second and delay in milliseconds. Reliability is how often it fails, how long recovery takes, and how the network holds up in a catastrophe. Forouzan lists the first three and cost is the fourth every real deployment adds — and it is the one that settles the argument, because the other three all get bought with it.
Throughput and delay are both performance. What is the difference?
Throughput is how many bits per second the link moves; delay is how long one message takes to get there, and they are different numbers that can move in opposite directions. Pushing more traffic through a link raises throughput and raises queueing delay at the same time. Delay itself splits into transmission time, which depends on the link rate, propagation time, which depends on distance and the speed of the signal in the medium, and queueing time at each device on the path, which is the piece that grows as you load the link harder.
Simplex, half duplex and full duplex, with an example of each.
Simplex carries one direction permanently — a keyboard to its computer. Half duplex carries both directions but only one at a time — a walkie-talkie. Full duplex carries both at once — a telephone call, where both people can talk and be heard together. Give the direction rule and then the example; a definition on its own invites the follow-up.
Why does half duplex need a turnaround delay and full duplex does not?
Half duplex has one channel that both ends share, so before B may transmit, A has to stop driving the line and the electronics at both ends have to swap transmitter for receiver. That reversal takes time and carries no data. Full duplex has two independent channels, so neither end ever waits for the other to release anything, and there is nothing to reverse.
Is Wi-Fi full duplex?
No. A Wi-Fi station cannot transmit and receive on the same channel at the same instant — its own transmitter would drown the incoming signal — so 802.11 is half duplex on the radio and uses CSMA/CA to decide whose turn it is. A wired switched Ethernet link is genuinely full duplex, which is one concrete reason the cable feels faster than the air.
What does a network buy an organisation that justifies the cost?
Resource sharing, communication, centralised administration and backup, reliability through redundancy, and scalability. The answer that lands names the cost beside each one: a shared printer is also a queue and a single point of failure, and redundancy means paying twice for capacity you use once. A use with no cost attached sounds like a brochure.
Who actually writes these standards?
Two families. The IEEE standardises what happens on the wire and in the air — 802.3 is Ethernet, 802.11 is Wi-Fi. The IETF standardises what runs on top and publishes each specification as an RFC — IPv4 is RFC 791, and TCP was RFC 793 before being rewritten as RFC 9293. Naming a protocol with its number costs one extra second and is worth it.

08 Practice problems

Six to work through

For the ones with numbers in them, write the units down beside every figure before you start. Bits and bytes, and megabits and megabytes, are where these go wrong, and they go wrong quietly.

Label the rooftop sensor

Easy
A weather sensor on a rooftop sends a temperature reading over a radio link to a logger in the basement, which writes it to a file. Name all five components of that data communication, say which one the sentence above does not give you, and name the transmission mode of the link with a one-line reason.
Follow-up
Four of the five components are stated outright in one sentence and the fifth is not mentioned at all — which is exactly the one that gets left out of interview answers, so the sentence is a fair model of the mistake.
Show the hint
Two questions settle the whole thing: what would the logger already have to know before the first bit arrives, and does anything at all ever travel from the basement to the roof?

Sort the brochure claims

Easy
A vendor brochure makes four claims: 99.99% uptime; 10 Gbps per port; AES-256 encryption on every link; under 5 ms to any campus building. Name the criterion each claim is quoting, and name the one criterion that none of the four claims mentions.
Follow-up
Two of the four claims are the same criterion measured in two different units, so you will end up with fewer distinct criteria than claims. Finding the missing one therefore means counting the full list from memory, not reading the brochure harder.
Show the hint
Write the question each claim answers before you write any criterion name, then count how many distinct questions you actually ended up with.

Cost the turnaround

Medium
A half-duplex link runs at 1 Mbps and takes 10 ms to turn the line around. Device A sends a 5,000-bit request and device B sends a 5,000-bit reply that cannot begin until the request has fully arrived. Ignoring propagation delay, give the total time from A’s first bit to B’s last bit, give the same total on a full-duplex link running 1 Mbps in each direction, and state how many milliseconds of the half-duplex run carried no data.
Follow-up
The reply cannot start until the request has finished, so full duplex overlaps nothing here, unlike the console’s run where both ends had data ready. The saving is the turnaround and nothing else, which means how big a win full duplex is depends entirely on how the turnaround compares with the message.
Show the hint
Convert 5,000 bits at 1 Mbps into milliseconds first, then write the three durations of the half-duplex run in order before you add anything.

Choose the mode and defend it

Medium
For each of these three links, choose simplex, half duplex or full duplex and justify the choice in one sentence: a fire alarm panel reporting to a monitoring centre; two air-traffic controllers sharing one radio frequency; a lift emergency intercom on which the trapped passenger and the control room have to be able to interrupt each other.
Follow-up
One of the three can legitimately be built either of two ways, and the deciding argument is not a definition at all — it is the fourth criterion arguing with the second. Say which way you would build it and what the extra money buys.
Show the hint
For each link ask two questions in order: does anything ever need to travel the other way, and if it does, does it ever need to travel at the same time?

Find the break-even message size

Medium
A server sits 300 km away over fibre in which the signal travels about 200,000 km per second. Working on a 100 Mbps link, find the one-way message size, in bits, at which transmission time first equals propagation delay, convert it to bytes, and say in one sentence what that size tells you about when upgrading the link to 1 Gbps is worth paying for.
Follow-up
The two delays are made of different ingredients: one of them changes when you buy a faster link and the other does not change at all. The break-even is therefore a size in bits, not a matter of opinion, and below it the faster link buys nothing a user can feel.
Show the hint
Set the two delays equal and solve for the number of bits. The distance and the signal speed give you one side of that equation; the link rate gives you the other.

Quote the campus link

Hard
Two campus buildings 800 m apart must be connected. The written requirement is at least 500 Mbps of sustained throughput one way, no more than 4 hours of total downtime per year, and the link will carry student records. Three quotes arrive: (a) one fibre pair on one route, 1 Gbps full duplex, and the last fibre cut on campus took 26 hours to repair; (b) two fibre pairs on separate routes, 1 Gbps full duplex each, at 2.4 times the cost of (a); (c) a wireless bridge advertising 1 Gbps, half duplex on one shared channel, measured at 420 Mbps sustained one way. For each quote say whether it meets the written requirement, naming the criterion it breaks and the exact number that breaks it wherever it does not, then say which you would buy and what the extra money is buying, and name the one criterion that none of the three quotes addresses at all.
Follow-up
Every quote advertises 1 Gbps, so the headline number eliminates nobody — you have to reach past it to a measured figure and to a figure about time rather than speed. And one of the four criteria is not answered by any of the three quotes, which means it is still your problem after you have signed.
Show the hint
Compare a single repair incident against the whole yearly downtime allowance, and the one measured throughput figure against the one required figure. Then go back to the list of four criteria and ask which one no quote mentions.