Networking Basics

Why start here

A computer network is a set of devices that can exchange data with each other over some shared medium: copper cable, optical fibre, or radio waves. Every app you build talks over one. When a page loads slowly, a video buffers, or a database call times out, the cause is usually one of a handful of basic quantities: how fast bits are pushed onto a wire, how long they take to travel, and how long they wait in queues along the way.

This lesson builds that vocabulary from zero. You will learn the kinds of networks (LAN, WAN and friends), the shapes they come in (topologies), the two ways of moving data across them (circuit and packet switching), and the four delays every packet suffers. Then you will put numbers on all of it: throughput, bandwidth, latency, round-trip time and the bandwidth-delay product.

Interviewers for freshers commonly probe three things here. First, definitions: "What is the difference between bandwidth and throughput?" Second, small calculations: "How long does it take to send a 1,500-byte packet over a 10 Mbps link 2,000 km long?" Third, reasoning: "Why did the internet choose packet switching over circuit switching?" By the end you should be able to answer all three without hesitation.

What a network is made of

Before types and shapes, it helps to name the parts.

• Host (end system): a device that runs applications and is the source or final destination of data. Your laptop, a phone, a web server and a smart TV are all hosts.
• Link: the physical connection between two devices. It has a rate (how many bits per second it can carry) and a length (which decides how long a bit takes to travel).
• Packet switch: a device in the middle that receives data on one link and forwards it out on another. The two main kinds are switches, which forward inside a local network using hardware (MAC) addresses, and routers, which forward between networks using IP addresses.
• Protocol: an agreed set of rules for the format and order of messages and the actions taken when a message arrives. HTTP, TCP, IP and Ethernet are protocols.
• Packet: a chunk of data plus a header. The header carries addressing and control information, much like the address written on an envelope.

 +--------+      +--------+      +--------+      +--------+
 | Host A |----->| Switch |----->| Router |----->| Host B |
 +--------+ link +--------+ link +--------+ link +--------+
   source          local          between         destination
                   forwarding     networks

The internet is simply a very large "network of networks": millions of local networks joined by routers that belong to internet service providers (ISPs). No single organisation owns it; protocols are what make the pieces cooperate.

Types of networks by size

Networks are often classified by how much geographic area they cover. The boundaries are fuzzy, so treat these as typical rather than strict.

PAN: personal area network

A personal area network (PAN) connects devices around one person, usually within a few metres. Your phone talking to wireless earbuds over Bluetooth, or a smartwatch syncing with a phone, is a PAN. Data rates are modest and power use is kept low because the devices run on batteries.

LAN: local area network

A local area network (LAN) covers a single building or campus: a home, an office floor, a college hostel. It is owned and managed by one organisation. LANs are fast (1 Gbps on ordinary office Ethernet, 10 Gbps or more in data centres), have low delay and few errors. Ethernet (wired) and Wi-Fi (wireless) are the two dominant LAN technologies.

MAN: metropolitan area network

A metropolitan area network (MAN) spans a city. A cable-TV operator's network across a city, or a fibre ring linking a university's several campuses in the same city, are examples. It is bigger than a LAN but usually still run by one provider.

WAN: wide area network

A wide area network (WAN) spans countries or continents. A company linking its Bengaluru, Pune and London offices uses a WAN, often leased from telecom carriers. WAN links are long, so propagation delay is significant, and bandwidth costs much more per bit than on a LAN. The internet itself is the largest WAN.

Network topologies

A topology is the shape of the connections between devices. There are two views. The physical topology is how cables actually run. The logical topology is how data actually flows. A classic example of them differing: old Ethernet hubs were wired as a star (physical) but behaved like a shared bus (logical), because the hub repeated every signal to every port.

Bus

All devices attach to a single shared cable called the backbone. A signal sent by one device travels along the whole cable and every device sees it.

  [A]     [B]     [C]     [D]
   |       |       |       |
 ==+=======+=======+=======+==   <- one shared cable (terminated)

Cheap and simple, but one break in the backbone takes down everything, and only one device can transmit at a time without collisions. Early Ethernet (10BASE5 "thick" coax) was a bus.

Star

Every device connects to a central device: a hub or, today, a switch.

          [A]   [B]
            \   /
   [F] --- [Switch] --- [C]
            /   \
          [E]   [D]

A broken cable affects only one device, and adding devices is easy. The central device is a single point of failure. Almost every modern wired LAN is a physical star with a switch in the middle.

Ring

Each device connects to exactly two neighbours, forming a loop; data passes around the ring. Token Ring and FDDI used this. A single break can stop the whole ring unless a second, counter-rotating ring is added for redundancy (as FDDI did).

Mesh

In a full mesh, every device links directly to every other device. With n devices you need n(n-1)/2 links. For 6 devices that is 6 × 5 / 2 = 15 links; for 50 devices it is 1,225 links. Full mesh gives the most redundancy and no shared bottleneck, but cabling grows quadratically, so it is used only for small sets of critical nodes (for example, core routers). A partial mesh connects only some pairs and is how the internet backbone looks.

Tree and hybrid

A tree (hierarchical) topology is a star of stars: an access switch per floor connects up to a distribution switch, which connects up to a core switch. Most campus networks look like this. A hybrid topology mixes any of the above.

Circuit switching vs packet switching

There are two basic ways to move data through a network of links and switches.

Circuit switching

In circuit switching, the network reserves a dedicated path with a fixed rate for the whole conversation before any data is sent. The classic telephone network works this way. There are three phases: set up the circuit, transfer data, tear down the circuit.

A link is shared among circuits by dividing it either in frequency or in time:

• Frequency-division multiplexing (FDM): each circuit gets its own frequency band all the time (like radio stations).
• Time-division multiplexing (TDM): time is cut into repeating frames, and each circuit gets one fixed slot in every frame.

Worked example: circuit-switched file transfer. A link of 2.048 Mbps uses TDM with 32 slots per frame. Setting up a circuit takes 400 ms. How long does it take to send a file of 1,280,000 bits?

1. Each circuit gets 1/32 of the link: 2,048,000 / 32 = 64,000 bits per second (64 kbps).
2. Transfer time = 1,280,000 / 64,000 = 20 s.
3. Add setup: 20 + 0.4 = 20.4 s.

Notice that even if the other 31 slots are idle, this circuit cannot use them. That waste is the core weakness of circuit switching for bursty data.

Packet switching

In packet switching, data is broken into packets that are sent without reserving anything. Each packet carries the destination address in its header, and every switch along the way looks at it and forwards it. Links are shared on demand: whoever has packets uses the full link rate. This is called statistical multiplexing.

Most packet switches use store-and-forward: a switch must receive the whole packet before it starts sending it on the next link. If packets arrive faster than the outgoing link can send them, they wait in an output queue (buffer). If the buffer is full, newly arriving packets are dropped. That is packet loss.

Worked example: why packet switching wins for bursty users. A 1 Mbps link is shared by users who each need 100 kbps while active but are active only 10% of the time.

• With circuit switching, each user needs a reserved 100 kbps, so the link supports exactly 1,000 / 100 = 10 users.
• With packet switching, suppose 35 users share it. Trouble occurs only when more than 10 are active at the same moment. The number active follows a binomial distribution with n = 35 and p = 0.1. The probability that 11 or more are active is about 0.0004.

So packet switching serves 3.5 times as many users, and 99.96% of the time they get the same performance as with circuits. This is the main reason the internet is packet-switched.

from math import comb
p = sum(comb(35, k) * 0.1**k * 0.9**(35 - k) for k in range(11, 36))
print(round(p, 6))   # 0.000424

There is also message switching, an older store-and-forward scheme where the whole message (not cut into packets) is stored at each node. It needs large buffers and adds long delays, so it is mostly of historical interest. Packet switching is message switching with the message cut into small pieces, which lets different pieces be on different links at the same time (pipelining).

Unicast, multicast and broadcast

These words describe how many receivers a single sent packet is meant for.

• Unicast: one sender to exactly one receiver. Loading a web page is unicast. Almost all internet traffic is unicast.
• Broadcast: one sender to every device on the local network. ARP requests ("who has IP 192.168.1.1?") and DHCP discovery are broadcasts. Routers do not forward broadcasts between networks, which keeps them from flooding the internet. The Ethernet broadcast address is ff:ff:ff:ff:ff:ff; the IPv4 "limited broadcast" address is 255.255.255.255.
• Multicast: one sender to a chosen group of receivers that have joined that group. IPv4 multicast addresses are in 224.0.0.0/4 (224.0.0.0 to 239.255.255.255). IPTV inside an ISP and routing protocols such as OSPF (which uses 224.0.0.5) use multicast. Multicast is rarely routed across the public internet.
• Anycast: one sender to the nearest of several receivers that share the same address. Public DNS resolvers and CDNs use anycast so you reach the closest server.

Unicast      A ---> B

Broadcast    A ---> B, C, D, E      (everyone on the LAN)

Multicast    A ---> B, D            (only members of the group)

Anycast      A ---> nearest of {S1, S2, S3} sharing one address

The four delays

When a packet travels from one node to the next, its nodal delay is the sum of four parts:

d_nodal = d_proc + d_queue + d_trans + d_prop

Picture one packet arriving at router A and leaving towards router B:

 packet    +-------------------- Router A ---------------+
 arrives   | check   | wait in      | push bits  |        travel
 ------->  | header  | output queue | onto link  | -----> along link
           | d_proc  | d_queue      | d_trans    |        d_prop
           +---------------------------------------------+   to B

1. Processing delay

Processing delay is the time a router takes to examine the header, check for bit errors and decide which outgoing link to use. In modern high-speed routers it is in microseconds or less.

2. Queuing delay

Queuing delay is the time a packet waits in the output buffer before its turn to be transmitted. It depends entirely on how many packets are ahead of it, so it changes from packet to packet. It can be zero on an idle link or many milliseconds on a congested one.

A useful rule of thumb uses traffic intensity, La/R, where L is packet size in bits, a is the average packet arrival rate in packets per second, and R is the link rate in bits per second.

• If La/R is close to 0, packets rarely wait.
• As La/R approaches 1, average queuing delay grows very quickly (not linearly).
• If La/R is greater than 1, bits arrive faster than they can leave; the queue grows without bound until packets are dropped.

Example: 1,500-byte packets (12,000 bits) arrive at 500 per second on a 10 Mbps link. Traffic intensity = 12,000 × 500 / 10,000,000 = 0.6. The link is busy 60% of the time; queuing is noticeable but stable.

3. Transmission delay

Transmission delay is the time to push all bits of the packet onto the link. If the packet has L bits and the link rate is R bits per second:

d_trans = L / R

It depends on packet size and link rate, not on distance.

4. Propagation delay

Propagation delay is the time for one bit to travel from one end of the link to the other. If the link length is d and the signal speed is s:

d_prop = d / s

Signals in copper and fibre travel at roughly 2 × 10^8 m/s (about two-thirds of the speed of light in vacuum); radio in air is close to 3 × 10^8 m/s. Propagation delay depends on distance, not on packet size or link rate.

Worked example 1: one link

A 1,500-byte packet is sent over a 10 Mbps link that is 2,000 km long. Signal speed is 2 × 10^8 m/s. Processing takes 20 µs and the packet waits 0.5 ms in the queue. Find the total delay.

1. Packet size in bits: L = 1,500 × 8 = 12,000 bits.
2. Transmission: 12,000 / 10,000,000 = 0.0012 s = 1.2 ms.
3. Propagation: 2,000,000 m / 200,000,000 m/s = 0.01 s = 10 ms.
4. Queuing: 0.5 ms. Processing: 20 µs = 0.02 ms.
5. Total: 1.2 + 10 + 0.5 + 0.02 = 11.72 ms.

On a long link, propagation dominates. Upgrading the link to 1 Gbps would shrink transmission to 0.012 ms but leave the 10 ms propagation untouched. Bandwidth does not make light go faster.

Worked example 2: several links, several packets

Host A sends to host B through two routers, so there are 3 links, each 10 Mbps with 1 ms propagation delay. Ignore processing and queuing. Packets are 12,000 bits.

One packet. With store-and-forward, each router waits for the full packet before forwarding it, so the packet pays transmission delay on each link:

end-to-end = N × (L/R) + N × d_prop
           = 3 × 1.2 ms + 3 × 1 ms = 6.6 ms

Five packets back to back. Packets pipeline: while packet 1 is on link 2, packet 2 is on link 1. The last packet leaves A after 5 transmission times and then needs N - 1 more transmissions to cross the remaining links:

end-to-end = (N + P - 1) × (L/R) + N × d_prop
           = (3 + 5 - 1) × 1.2 ms + 3 × 1 ms
           = 8.4 ms + 3 ms = 11.4 ms

Timeline of transmissions (each cell is 1.2 ms, numbers are packet ids):

time slot:  1   2   3   4   5   6   7
link 1:     1   2   3   4   5
link 2:         1   2   3   4   5
link 3:             1   2   3   4   5

The last packet finishes on link 3 at slot 7, that is 7 × 1.2 = 8.4 ms of transmission, plus the propagation along three links.

Worked example 3: when is transmission bigger than propagation?

Compare transmission and propagation for a 1,000-bit frame on a 100 m LAN cable at 1 Gbps.

• Transmission: 1,000 / 10^9 = 1 µs.
• Propagation: 100 / (2 × 10^8) = 0.5 µs.

Here transmission is twice propagation. On short LANs the two are comparable; on long WAN links propagation usually dominates. The ratio a = d_prop / d_trans reappears in the data link and transport lessons, where it decides how efficient stop-and-wait is.

Throughput, bandwidth and latency

These three words are often mixed up. Pin them down.

• Bandwidth (strictly, link rate or capacity): the maximum number of bits per second a link can carry. A "100 Mbps link" has bandwidth 100 Mbps. In signal processing, bandwidth means a frequency range in hertz; in networking interviews it almost always means bit rate.
• Throughput: the rate at which data is actually delivered, end to end, in bits per second. It is never more than bandwidth and is often much less because of sharing, protocol overhead, loss and waiting for acknowledgements.
• Latency: how long it takes data to get from source to destination. For one packet, it is the sum of all nodal delays along the path. People often say latency to mean one-way delay.
• Round-trip time (RTT): the time for a small packet to go from client to server and for the reply to come back. ping reports RTT. Most protocol costs are counted in RTTs: a TCP handshake costs one RTT before data flows, and TLS 1.3 adds one more.
• Jitter: the variation in delay from packet to packet, caused mainly by queuing. It matters for voice and video calls.

Bottleneck throughput

Data from a server reaches you over a chain of links. The end-to-end throughput is limited by the slowest link on the path, the bottleneck link:

throughput = min(R1, R2, ..., Rn)

Example: server uplink 100 Mbps, an ISP core link with a fair share of 50 Mbps, your home broadband 20 Mbps. Throughput is min(100, 50, 20) = 20 Mbps. Upgrading the server to 1 Gbps changes nothing for you.

Time to download a 10 MB file (10 × 10^6 bytes) at 100 Mbps, ignoring everything else: 10 × 10^6 × 8 / 10^8 = 0.8 s.

The bandwidth-delay product

The bandwidth-delay product (BDP) is bandwidth multiplied by delay. It answers: how many bits are "in flight" in the pipe at once? Usually the delay used is RTT, because a sender must keep sending until the first acknowledgement returns.

BDP = bandwidth × RTT

Picture the link as a pipe: bandwidth is its cross-section, delay is its length, and BDP is its volume.

Worked example: BDP

A path has bandwidth 100 Mbps and RTT 40 ms.

1. BDP = 100 × 10^6 bits/s × 0.040 s = 4,000,000 bits.
2. In bytes: 4,000,000 / 8 = 500,000 bytes, about 500 KB.

Meaning: to keep this path fully busy, a sender must have about 500 KB of data sent but not yet acknowledged at all times. If its window is smaller, the pipe sits partly empty.

Worked example: a small window wastes a fast link

Classic TCP without window scaling allows at most 65,535 bytes unacknowledged. Over a path with RTT 100 ms, the sender can send one window per RTT at most:

max throughput = window / RTT
               = 65,535 × 8 bits / 0.1 s
               ≈ 5.24 Mbps

Even if the link is 1 Gbps, the sender cannot exceed about 5.24 Mbps. This is why TCP added the window scale option, which you will meet in the TCP flow and congestion control lesson. Networks with a large BDP (fast and far, such as satellite or intercontinental fibre) are called long fat networks.

A satellite example

A geostationary satellite orbits about 35,786 km above the equator. A signal going ground to satellite takes about 35,786,000 / (3 × 10^8) ≈ 119 ms, and ground to satellite to ground about 239 ms. A request and its reply need that twice, so RTT is close to half a second before any processing. No bandwidth upgrade can fix that; it is why low-earth-orbit constellations, at a few hundred kilometres up, have much lower latency.

Client-server vs peer-to-peer

The last basic idea is how applications are organised across hosts.

Client-server

In the client-server model, an always-on host called the server provides a service, and clients connect to it to ask for it. Clients do not talk to each other directly. The server has a fixed, well-known address (usually reached through a DNS name). The web, email and most mobile app backends are client-server.

   [Client]   [Client]   [Client]
        \         |         /
         \        |        /
          +---> [Server] <+
                (always on, known address)

Strengths: one place to manage data, security and updates; easy for clients to find. Weaknesses: the server is a bottleneck and a single point of failure, so large services spread the load over many servers in data centres (see load balancing).

Peer-to-peer

In peer-to-peer (P2P), hosts called peers talk directly to each other, and each peer can act as both client and server. BitTorrent file sharing is the classic example; WebRTC video calls send media peer-to-peer when they can.

   [Peer] <-----> [Peer]
     ^  \        /  ^
     |   \      /   |
     v    v    v    v
   [Peer] <-----> [Peer]

Strengths: self-scalability (every new peer brings upload capacity as well as demand) and no single point of failure. Weaknesses: peers come and go, they sit behind NAT and firewalls, and security and trust are harder.

Worked example: distributing a file

Distribute a file of size F bits to N hosts. The server's upload rate is us; each peer's download rate is at least dmin; each peer uploads at u.

• Client-server: the server must upload N copies, so time is at least max(N·F/us, F/dmin). It grows linearly with N.
• Peer-to-peer: the server need upload only one copy, and peers redistribute. The minimum time is at least max(F/us, F/dmin, N·F/(us + N·u)). As N grows, the last term approaches F/u and levels off.

With F = 1 Gbit, us = 100 Mbps, u = 10 Mbps, dmin large, and N = 1,000:

• Client-server: 1,000 × 1 / 0.1 = 10,000 s.
• Peer-to-peer lower bound: 1,000 × 1 / (0.1 + 1,000 × 0.01) = 1,000 / 10.1 ≈ 99 s.

Many real systems are hybrids: a central server for login and discovery, then direct peer connections for heavy data.

Putting it together: what happens to one packet

Walk one packet from your laptop to a server in another city to see how the ideas fit.

1. Your app hands data to the operating system, which wraps it in headers and puts it in a packet (packet switching: no reservation).
2. The laptop's Wi-Fi card transmits it (transmission delay, L/R of the Wi-Fi link) to the home router across a few metres (tiny propagation delay). This is your LAN.
3. The home router examines the header (processing delay), queues it behind your sibling's video stream (queuing delay), and sends it on the broadband link to the ISP.
4. ISP routers forward it across their WAN, each adding the four delays. The longest single term is usually propagation over hundreds of kilometres of fibre.
5. The server receives it, and the reply makes the reverse trip. The total is one RTT.
6. If the server sends a large file, your throughput is the rate of the bottleneck link, and the sender needs at least one bandwidth-delay product of data in flight to fill it.

The next lessons zoom into each step: how layers divide the work (OSI and TCP/IP), how frames cross one link (data link layer), how addresses and routes get the packet across networks, and how TCP makes the result reliable.

Interview questions

Q1. What is the difference between bandwidth and throughput?

Bandwidth is the maximum rate a link can carry, a property of the link. Throughput is the rate at which useful data is actually delivered end to end. Throughput is at most the bottleneck link's bandwidth and is usually lower because of sharing, protocol overhead, loss and waiting for acknowledgements.

Q2. Name the four components of nodal delay. Which ones depend on packet size?

Processing, queuing, transmission and propagation. Transmission delay (L/R) depends directly on packet size. Queuing depends on the sizes of packets ahead in the queue. Propagation depends only on distance and medium, and processing on the router's work per packet.

Q3. A 1,000-byte packet goes over a 1 Mbps link 3,000 km long (signal speed 2 × 10^8 m/s). Find transmission and propagation delay.

Transmission = 8,000 bits / 10^6 bit/s = 8 ms. Propagation = 3 × 10^6 m / 2 × 10^8 m/s = 15 ms. Ignoring queuing and processing, the last bit arrives after 23 ms.

Q4. Why does upgrading a link from 100 Mbps to 1 Gbps not reduce ping time much on an intercontinental path?

Ping uses tiny packets, so transmission delay is already negligible. RTT on long paths is dominated by propagation delay, which depends on distance and the speed of light in fibre. Faster links raise throughput for big transfers but do not shorten the distance.

Q5. What is the bandwidth-delay product and why does it matter?

It is bandwidth times delay (usually RTT): the amount of data that can be in flight on the path. A sender must be allowed to have at least that much unacknowledged data to keep the link full. It is used to size TCP windows and router buffers.

Q6. Why did the internet adopt packet switching instead of circuit switching?

Computer traffic is bursty: users send in short bursts and are idle most of the time. Circuit switching reserves capacity that sits unused during idle periods, while packet switching shares capacity on demand (statistical multiplexing) and supports many more users on the same link. The trade-off is variable delay and possible loss, which higher layers handle.

Q7. What is store-and-forward transmission?

A switch receives the whole packet, stores it, and only then begins transmitting it on the outgoing link. It allows error checking and makes forwarding decisions simple. Its cost is that each hop adds a full transmission delay, so a P-packet message over N links takes (N + P - 1) × L/R plus propagation.

Q8. What happens when traffic intensity La/R exceeds 1?

Bits arrive at the queue faster than the link can send them, so the queue grows without bound. In practice the buffer fills and new packets are dropped. Even below 1, delay grows sharply as intensity gets close to 1.

Q9. Differentiate unicast, multicast, broadcast and anycast.

Unicast is one sender to one receiver. Broadcast is one sender to all hosts on a local network. Multicast is one sender to a group that has joined. Anycast is delivery to the nearest of several hosts sharing one address, used by DNS resolvers and CDNs.

Q10. Compare star and mesh topologies.

A star connects every device to a central switch: easy to manage, a cable fault affects one device, but the switch is a single point of failure. A full mesh connects every pair: n(n-1)/2 links and maximum redundancy, but the cabling cost grows quadratically. Real networks use a star or tree at the edge and a partial mesh in the core.

Q11. What is the difference between latency and RTT?

Latency is usually one-way delay from sender to receiver. RTT is the time for a message to go and its reply to come back, so on a symmetric path it is about twice the one-way latency plus server processing. Protocol costs, such as a TCP handshake, are measured in RTTs.

Q12. Is P2P always faster than client-server for file distribution?

Not always. For a few receivers, a fast server can be quicker. P2P shines as the number of receivers grows, because every new peer adds upload capacity, so distribution time levels off instead of growing linearly. It also depends on peers actually uploading and on reachability through NAT.

Q13. How many links does a full mesh of 10 routers need, and how many ports per router?

10 × 9 / 2 = 45 links, and each router needs 9 ports. This growth is why full mesh is limited to small sets of critical devices.

Q14. What is jitter and which applications care about it?

Jitter is variation in packet delay, mostly from changing queuing delay. Real-time voice and video care because playback needs packets at a steady pace; they use a small playout buffer to smooth jitter at the cost of slightly higher latency. File downloads barely notice it.

Key takeaways

• A network is hosts, links, switches and routers, cooperating through protocols. The internet is a network of networks.
• PAN, LAN, MAN and WAN differ in reach and ownership; the internet is a WAN built from many smaller networks.
• Topologies: bus, star, ring, mesh (n(n-1)/2 links), tree. Modern LANs are physical stars; cores are partial meshes.
• Circuit switching reserves capacity; packet switching shares it on demand and wins for bursty data at the cost of queuing and loss.
• Nodal delay = processing + queuing + transmission (L/R) + propagation (d/s). Transmission depends on size and rate; propagation on distance.
• For P packets over N store-and-forward links: (N + P - 1) × L/R plus N propagation delays.
• Throughput is limited by the bottleneck link; bandwidth is capacity, not what you get.
• BDP = bandwidth × RTT is the data in flight needed to fill a path; a 64 KB window over 100 ms caps throughput near 5.24 Mbps.
• Always convert bytes to bits and use powers of ten for rates.

Next lesson

Continue with the OSI and TCP/IP models.