How to use this reader

• Chapters 1 to 8 are the study content: each topic explained in plain language, with drawn diagrams, an example to remember it by, and at the foot of each card every question the papers set on it, word for word, with the part the card answers lit. The copy button on a card copies it to paste into Claude and ask about.
• Theory answers give the exam answer to every question that asks what a thing is, why it matters or how two things compare, at the length its marks deserve.
• Practical answers give the ones that ask how a thing works, step by step: CSMA/CD, the three-way handshake, sliding window, DNS resolution and the like.
• Numericals work every calculation the papers set (subnetting, CRC, Hamming code, delays, efficiency, routing tables, RSA, the buckets), grouped by method, with the method once at the top of each group.
• Summary gives every topic as the skeleton of its answer, for the last read before the exam, and ends with the Recall sheet: every topic again as bare keywords.
• Compact chapters keep every detail of a chapter with the teaching prose taken out; one button copies a whole chapter to give Claude as context.
• Question bank reproduces all 269 questions word for word, by paper or by chapter, each linked to its answer and its card.
• Mind map draws each chapter as its lists; Close all turns it into a test. Flashcards drill the definitions and every question.

Writing the paper

• About two minutes a mark: 180 minutes for 80 marks, so an 8 mark question gets about 16 minutes.
• Most questions come in parts (2+6, 4+4, 2+3+3). Answer the parts in the order asked and give each its share of the time: a 2 mark definition is two or three lines.
• Draw what has a drawing: the layer models, the header and frame formats, the CSMA/CD flowchart, the handshake, the switching and tunnelling pictures each earn marks.
• Numericals show their working: the given values, the formula or the method, every step, then the answer in bold. A subnet plan is a table; a CRC is the long division.
• The last question is short notes, any two of four, about 4 marks each: a definition, three or four points and a small diagram.

The 27 sittings, question by question

Each paper opens to its questions, the chapter each belongs to and the card that teaches it. The board repeats itself: OSI and TCP/IP, CRC, subnetting, CSMA/CD, routing, TCP and UDP, DNS and IPv6 come back paper after paper.

The whole subject on one page

Eight chapters and every topic card in them. The number beside a topic is how many of the 27 sittings asked it.

How to read the chips

Under each chip is the list of sittings that asked it: 81 Bh is the 2081 Bhadra regular paper (bold, a regular sitting); 81 Ba is the 2081 Baishakh back paper.

What a computer network is, and what it is used for PIN 4/27

81 Bh · 72 Ch · 71 Shr · 66 Po2+65+3

Autonomous is the word to keep. Each computer can work on its own; the network only lets them talk. A mainframe with dumb terminals is not a network in this sense, because a terminal can do nothing alone. Two computers are interconnected when they can exchange information, whatever the medium between them.

Every network has four parts:

• Nodes: end systems (hosts) that run the applications, such as laptops, phones and servers, and the intermediate devices that move data between them: switches, routers and access points (devices).
• Links: the transmission media: twisted pair, coaxial cable, optical fibre, radio (media).
• Protocols: the rules both ends follow to understand each other (protocols).
• Services: what the users finally get: the web, mail, file sharing, voice and video calls.

Three criteria judge a network: performance (throughput and delay, see chapter 2), reliability (how often it fails and how quickly it recovers) and security (protection of the data against unauthorised access and damage, chapter 8).

The uses of computer networks, under the book's three headings, with the effect on society added:

Five instances of networks in a student's day (the 2072 Chaitra paper asks for exactly five):

1. Paying by phone: scanning a fonepay QR code at the canteen or loading an eSewa or Khalti wallet: the app is a client talking to the bank's server over mobile data.
2. Calling family: a WhatsApp or Viber video call to a relative working abroad travels as IP packets (VoIP), for the price of the data instead of an international call.
3. Studying: online classes, notes shared in the class group, and exam results checked on the exam board's website.
4. Entertainment: a YouTube video streamed on demand, or an online multiplayer game whose moves must arrive within milliseconds.
5. Daily services: booking a Pathao or inDrive ride (the phone's GPS position sent over mobile data), or paying the NEA electricity bill online instead of queuing at the office.

To remember the idea, picture a hostel kitchen: every student can cook alone (autonomous), but sharing one gas cylinder and one fridge (resources) works only with agreed rules about turns and shelves (protocols). The hostel Wi-Fi is the same thing in network form: one fibre line from the ISP shared by every room.

Networks by size and geography: PAN, LAN, MAN and WAN PIN 1/27

70 Asa3+5

Distance decides almost everything else. Over a few metres a cheap radio is enough. Inside one building an organisation can lay its own cable, so a LAN is privately owned, fast and nearly error free. Across a country nobody lays private cable: the links are leased from telecom carriers, cost more per bit and add delay. So the size of a network fixes its owner, its speed, its delay and its error rate.

The WAN's structure. A WAN joins hosts through a communication subnet: switching elements (routers) joined by transmission lines. Packets travel store and forward: each router receives a whole packet, stores it, then sends it on along the next line (switching). The hosts belong to the users; the subnet usually belongs to a carrier or an ISP.

Two other ways to classify networks:

• By transmission technology: broadcast networks share one channel that every machine hears (classic Ethernet, Wi-Fi), so an address in each frame says whom it is for; point-to-point networks join pairs of machines, so a packet may cross several intermediate nodes on its way (most WANs).
• By architecture and shape: client/server or peer to peer (networking models); bus, star, ring or mesh (topologies).

To remember it, follow one video call outwards: earbuds paired with the phone (PAN), the phone on the hostel Wi-Fi (LAN), the hostel on the ISP's fibre ring across the valley (MAN), the ISP's links across the border to the rest of the world (WAN), and all of them together (the Internet).

Network topologies: bus, star, ring, mesh, tree and hybrid PIN 1/27

70 Asa3+5

Physical and logical can differ. A classic Ethernet hub is wired as a star but behaves as a bus, since every frame it receives is repeated out of every port; a Token Ring is wired as a star into a central access unit, yet the token travels round it as a ring.

The mesh formula. A full mesh of $n$ nodes needs one link for every pair, and every node needs $n-1$ ports:

To remember them, think of how a village gets its water: a bus is one pipe along the road with every house tapped onto it (cut the pipe and the whole lane is dry); a star is a tank with a separate pipe to each house; a ring is a loop main round the village; a mesh is a pipe from every house to every other house: nothing ever runs dry, and nobody can afford it.

Active networking, compared with the traditional legacy network PIN 1/27

66 Bh3+5

The legacy network is passive. A traditional router only stores and forwards: it reads the header, looks up the route and sends the packet on, never touching the payload. Its functions are fixed by the vendor, so a new network service (a new multicast or quality of service scheme) needs years of standardisation and then a firmware upgrade of every router. Active networking, proposed by Tennenhouse and Wetherall at MIT in the mid 1990s and funded by DARPA, attacks exactly that slowness: it puts the new service into the network as a program.

Two ways to get the code into the node:

• Discrete approach (programmable switches): programs are loaded into the nodes beforehand, out of band, by an operator or an authorised user; arriving packets carry only a header that says which program should process them.
• Integrated approach (capsules): every packet, called a capsule, carries a small program as well as data; each node it reaches executes that code, which decides what happens to the capsule. The MIT ANTS toolkit worked this way.

The framework of an active node (the DARPA active network architecture) has three layers of software on the node's hardware:

1. NodeOS: the node operating system. It owns the node's resources (the links, called channels, processor time, memory and storage), shares them among the execution environments, and enforces security so that no program takes more than its share.
2. Execution environments (EEs): each is like a virtual machine or interpreter (a Java virtual machine, for example) that runs active code. One node may host several, and a management EE lets the operator control the node.
3. Active applications (AAs): the user programs that run inside an EE and give a flow its custom service.

An arriving packet is matched to the right EE by a small header (the Active Network Encapsulation Protocol, ANEP), is processed there, and leaves possibly changed: forwarded, merged, shrunk, copied or dropped.

What it is good for: shrinking a video stream at the node nearest a slow link; caching popular content inside the network; merging the readings of thousands of sensors on the way instead of carrying them all; deploying a new multicast or congestion control scheme without waiting for a standard; pushing firewall rules to the right node during an attack; network management by mobile agents.

Where the idea went. Few active networks were deployed, because running other people's code inside routers raised hard security and performance problems. The idea of a programmable network survived in software-defined networking (SDN), where a central controller programs the switches, and in programmable switch hardware.

To remember it: a legacy network is the postal service, which reads only the address and passes the parcel on. An active network is a courier who also obeys a note on the parcel: "if the road to the village is slow, open me and send only the small photos".

Services: connection-oriented and connectionless, and the service primitives PIN 1/27

75 Ash2+4+2

Connection-oriented service is modelled on the telephone system: the user first establishes a connection, uses it to send data, then releases it. The connection acts like a tube: bits go in at one end and come out in the same order at the other. The two sides may negotiate parameters, such as the maximum message size or the quality of service, when the connection opens. Examples: TCP, X.25 virtual circuits, a phone call.

Connectionless service is modelled on the postal system: each message (datagram) carries the full destination address and is routed independently of the others, so two messages to the same place may take different routes and arrive out of order. Examples: UDP, IP.

Reliable means the receiver acknowledges every message, so the sender knows it arrived; the acknowledgements cost delay, which is why voice and video often prefer an unreliable service.

Service primitives are the operations a user process calls to use a service; in an operating system they are usually system calls. Five primitives are enough for a simple connection-oriented service:

How a client and a server use them (six packets in all):

1. LISTEN: the server calls LISTEN and blocks until a connection request arrives.
2. CONNECT: the client calls CONNECT, which sends a connection request packet (1) to the server; the client is suspended until there is a response.
3. Accepted: the server's operating system sees the request, unblocks the server and sends back a packet accepting the connection (2); the client is released and the connection is up.
4. RECEIVE: the server calls RECEIVE to wait for the first request.
5. SEND: the client SENDs its request (3) and calls RECEIVE to wait for the answer; the server processes the request and SENDs the reply (4).
6. DISCONNECT: the client calls DISCONNECT (5); the server answers with its own DISCONNECT (6), and the connection is released.

Later editions of Tanenbaum add a sixth primitive, ACCEPT, for step 3. The Berkeley socket calls are the same idea in code: listen(), connect(), accept(), send(), recv() and close() (sockets, socket programming).

The four classes of OSI primitive: request (a user asks for a service, as in CONNECT.request), indication (the peer is told of the event), response (the peer answers) and confirm (the first user learns the result). A confirmed service uses all four; an unconfirmed service only the request and the indication.

Services and protocols are not the same. A service is what a layer does for the layer above (vertical: the operations it offers, not how they work); a protocol is the set of rules that peers on different machines use to carry out that service (horizontal: the format and meaning of the packets). A layer can change its protocol freely as long as the service it offers stays the same, much as a program keeps calling one function while the function's code is rewritten.

To remember it: a phone call is connection-oriented (dial, talk, hang up; the words arrive in order); a letter is connectionless (each envelope carries the full address, and two letters posted together may arrive on different days).

The TCP/IP model: four layers and their protocols PIN 4/27

82 Ba · 81 Ba · 80 Ba · 70 Asa4+41+3+43+5

Where it came from. TCP/IP grew out of the ARPANET, the research network funded by the US Department of Defense's Advanced Research Projects Agency (ARPA). Vinton Cerf and Robert Kahn described TCP in 1974; the ARPANET switched to TCP/IP on 1 January 1983, and the Internet grew from it. Its main goal was to interconnect many different networks and give universal communication services over them: connections had to survive the loss of routers and lines in between as long as the two ends kept working, and the design had to carry very different applications, from file transfer to real-time speech. The protocols came first; the model was written afterwards to describe them.

The four layers, bottom up:

1. Host-to-network (network access, link) layer: the lowest layer. The original model says little more than that the host must connect to the network with some protocol so that it can send IP packets over it. In practice it covers framing, physical (MAC) addressing and putting the bits on the medium: OSI's data link and physical layers together. Protocols: Ethernet (IEEE 802.3), Wi-Fi (IEEE 802.11), PPP, DSL, Frame Relay, ATM.
2. Internet layer: the linchpin that holds the whole architecture together. It lets a host inject packets into any network and have them travel independently to the destination, possibly by different routes and out of order: a connectionless, best effort service, like letters in the post. It defines an official packet format and protocol, IP, and handles logical addressing and routing. Protocols: IP (IPv4, RFC 791; IPv6, RFC 8200), ICMP for error and control messages, IGMP for multicast groups, ARP to find a MAC address (often placed at the boundary with the layer below).
3. Transport layer: lets peer processes on the source and destination hosts carry on a conversation, as OSI's transport layer does. It segments the data, reassembles it, and names the process by a port number. Two protocols: TCP, reliable and connection-oriented, which delivers a byte stream without error and in order, with flow control (used where accuracy matters: the web, mail, file transfer); and UDP, unreliable and connectionless, with no sequencing or flow control (used where prompt delivery matters more: DNS lookups, voice and video calls, online games).
4. Application layer: all the higher-level protocols users work with. TCP/IP has no session or presentation layer: applications include those functions themselves when they need them. Protocols: HTTP and HTTPS (web), SMTP, POP3 and IMAP (mail), FTP (files), DNS (names), DHCP (address assignment), SNMP (management), Telnet and SSH (remote login).

The protocols at each layer, with the well known port numbers of the application protocols (ports):

The hourglass. Many applications sit on two transport protocols, which sit on one internet protocol, which runs over every kind of link. IP is the narrow waist: anything that can carry IP packets can join the Internet, and any application written for IP works over any link. That one design choice is why the same browser works on the hostel Wi-Fi, a fibre line at home and a 4G phone.

Four layers or five? Tanenbaum's TCP/IP model, which the book follows, has four layers. Kurose and Ross, and Tanenbaum's own "hybrid model" for teaching, use five: application, transport, network, data link and physical. Both are right, as long as the answer says which one it draws.

To remember it, follow one Messenger video call from the hostel: the app (application) hands its audio and video to UDP (transport), which hands it to IP with the server's address (internet), which rides Wi-Fi frames to the router and then the ISP's fibre (host-to-network). Four handovers, four layers.

Data encapsulation: how headers and trailers are added and removed PIN 3/27

76 Ash · 70 Ch · 66 Bh4+42+2+4

Headers and trailers. A header is control information placed in front of the data: addresses (MAC, IP, port), sequence and acknowledgement numbers, a length, a type, a time to live, a checksum. A trailer is control information placed after the data; in practice it is the data link layer's frame check sequence (FCS), a CRC computed over the whole frame, and sometimes an end marker. The trailer goes at the end because the CRC can be calculated while the frame is being sent, and appended last.

PDU and SDU. What a layer receives from the layer above is its service data unit (SDU); the SDU plus the layer's header (and trailer) is its protocol data unit (PDU), which becomes the SDU of the layer below. Each PDU has its own name:

At the sender, five steps, as the book counts them in the OSI model:

1. Data: the application, presentation and session layers create the data from the user's input (a request, a message, a file).
2. Segment: the transport layer cuts the data into pieces and adds a TCP or UDP header (source and destination ports, sequence number, checksum).
3. Packet: the network layer adds an IP header (source and destination IP addresses, time to live, protocol number).
4. Frame: the data link layer adds a frame header (destination and source MAC addresses, type) and the trailer (FCS).
5. Bits: the physical layer turns the frame into a stream of bits, as electrical, light or radio signals on the medium.

At the receiver, the same steps in reverse. The physical layer turns the signals back into bits. The data link layer checks the FCS (a damaged frame is discarded), checks that the destination MAC address is its own, strips the header and trailer, and passes the packet up. The network layer checks the destination IP address and removes the IP header. The transport layer uses the port number to find the right process, puts the segments in order and removes its header, and the application receives the original data. Each layer reads only the header its peer wrote: the headers are how peers talk (virtual communication).

At a router the frame is opened only up to the network layer: the router reads the IP header, chooses the next hop, and wraps the packet in a new frame for the next link, with new MAC addresses. The IP addresses stay the same end to end; the MAC addresses change at every hop.

To remember it, post a letter: the letter itself is the data; it is sealed in an envelope with the friend's name (transport); the post office puts it in a bag tagged with the destination district (network); the bag rides a truck with a trip sheet in front and a seal at the back that is checked on arrival (the data link header and trailer); and the truck drives on the road (physical). At the other end each office opens only its own wrapping.

The Internet: a network of networks

How it grew:

• 1969, the ARPANET: four nodes in the US (UCLA, SRI, UC Santa Barbara and the University of Utah), funded by ARPA: the first large packet switched network.
• 1974 to 1983, TCP/IP: Cerf and Kahn's internetworking protocol, which the ARPANET adopted on 1 January 1983.
• 1986 to 1995, NSFNET: the US National Science Foundation's backbone joined the universities; it was retired in 1995, when commercial ISPs took over the backbone.
• 1989 to 1991, the World Wide Web: Tim Berners-Lee at CERN created the web (HTTP, HTML, URLs), which brought the Internet to the public.

How it is built:

• End systems (hosts): phones, laptops and servers at the edge.
• Access networks: the last link to the user: DSL, cable, fibre to the home, Wi-Fi, 4G and 5G.
• ISPs in tiers: local access ISPs buy transit from national and international ISPs, which connect to global (tier 1) backbones; ISPs of similar size often exchange traffic with each other free (peering).
• Internet exchange points (IXPs): places where many ISPs connect to swap traffic directly. In Kathmandu the Nepal Internet Exchange (NPIX) lets Nepali ISPs hand local traffic to each other without sending it abroad and back.
• Content providers' networks: large companies run their own global networks and place caches inside ISPs, so a popular video often comes from a server in the same city.

Who runs what. The Internet has no central government; each network sets its own policies. Only the two main name spaces are coordinated centrally: IP addresses (allocated by IANA, under ICANN, to five regional registries, of which APNIC serves the Asia Pacific region, Nepal included) and the root of the Domain Name System (under ICANN), below which .np is Nepal's country code domain (DNS). Standards come from the IETF as RFCs, with the Internet Society (ISOC) behind them.

Internet, intranet, extranet: an internet (small i) is any set of networks joined by routers; the Internet (capital I) is the global one; an intranet is a private network inside one organisation built with the same TCP/IP technology (a college portal reachable only on campus); an extranet opens part of an intranet to partners (suppliers logging in to a company's order system).

To remember it: run tracert (Windows) or traceroute (Linux) to a foreign website from the hostel. The list of hops shows the trip from the Wi-Fi router to the ISP, across the border and on to the server's network: a dozen independently owned networks passing the same packets along.

ATM: fixed 53-byte cells, virtual paths and the adaptation layers PIN 3/27

81 Bh · 80 Bh · 66 Bh2×43+3

Asynchronous because a source sends a cell whenever it has data, not in a fixed time slot of its own as in synchronous TDM (multiplexing), so no slot is wasted on an idle source. Cells rather than frames, because small cells of one size can be switched in hardware at very high speed, and a voice cell never waits behind a 1500-byte data frame: the delay stays low and predictable.

Why 48 bytes: a compromise. The US wanted 64-byte payloads for efficiency, Europe 32 bytes for low voice delay, and the committee settled between them on 48. Filling 48 bytes with 64 kbps voice takes $48\times 8/64000=6$ ms.

The cell header at the user network interface (UNI):

Between two switches, at the network network interface (NNI), there is no GFC: its four bits extend the VPI to 12.

Virtual paths and virtual channels. A physical link (the transmission path) carries many virtual paths, and each virtual path bundles many virtual channels. A connection is named by the pair VPI/VCI on each link, which has local significance like a DLCI. A VP switch (cross-connect) switches whole paths and changes only the VPI, so thousands of channels are rerouted with one table entry; a VC switch changes both. As in X.25 and Frame Relay, connections may be permanent (PVCs) or switched (SVCs, set up with Q.2931 signalling).

The ATM reference model has three layers (and three planes: user, control and management):

The AAL types, one for each class of service:

Service categories (ATM Forum) state what each connection is promised: CBR (constant bit rate: voice, video), rt-VBR (real-time variable: compressed video), nrt-VBR (non real-time variable: data with delay bounds), ABR (available bit rate: data that adapts to what is free) and UBR (unspecified: best effort, like IP).

Advantages: very high speed; voice, video and data on one network with real quality of service; low, predictable delay; scales from the desktop to the backbone. Disadvantages: the 9.4% cell tax, complexity and cost; for data it lost to cheaper switched Ethernet and to IP over MPLS. It ran telephone and Internet backbones in the 1990s, and carried ADSL traffic between home modems and the exchange (xDSL).

To remember it: ATM is a supermarket that ships everything, rice, eggs or a television, in identical 53-litre crates on a conveyor: the crates move fast and never jam, though a lot of crate travels with every egg.

Ethernet as an example network

From a shared cable to a switched star. Ethernet was invented at Xerox PARC by Robert Metcalfe and David Boggs in 1973; DEC, Intel and Xerox published the DIX Ethernet specification (10 Mbps, 1980, revised as Ethernet II in 1982), and the IEEE standardised it as 802.3 in 1983. The first Ethernets were a coaxial bus that every station shared, taking turns by CSMA/CD (CSMA/CD). Today every station has its own full-duplex link to a switch: a star with no collisions at all, which is why modern Ethernet no longer needs CSMA/CD.

The name code reads speed, signalling and medium: 10BASE-T is 10 Mbps, baseband, twisted pair; 100BASE-TX is Fast Ethernet over two pairs of Category 5 cable; 1000BASE-T is gigabit over four pairs; 10GBASE-SR is 10 gigabit over short-reach multimode fibre.

What every version keeps is the frame (preamble, 48-bit destination and source MAC addresses, type or length, 46 to 1500 bytes of data, a 32-bit CRC), so a frame from an old card is still understood by a switch bought today. The frame, MAC addressing and the access rules are taught in chapter 3 (Ethernet in detail).

Why it won: it is cheap and simple, every new generation is backward compatible, its speed has been multiplied by ten again and again, and the switch replaced the shared bus without changing the frame. Carriers now sell Metro Ethernet and Carrier Ethernet as MAN and WAN services, which is why the book calls it a technology "used in LANs and MANs".

To remember it: the blue cable from the hostel router to a desktop is Ethernet: most likely 1000BASE-T on Category 5e or 6 cable, a star point to the switch inside the router, and the same frame format Ethernet has used since the 1980s.

VoIP: voice calls as IP packets

Circuit against packet. A PSTN call reserves a 64 kbps circuit end to end for the whole call, silences included (the telephone network); VoIP sends the voice as packets that share the network with everything else, which is why it avoids the tolls of the traditional network and costs only the data.

How a VoIP call works:

1. Signalling: a signalling protocol finds the other party and sets up the call: SIP (the Session Initiation Protocol, RFC 3261: INVITE, ringing, 200 OK, ACK, and BYE to hang up) or the older ITU-T H.323.
2. Digitise and compress: the microphone's signal is sampled (8000 times a second for ordinary telephone quality) and compressed by a codec: G.711 (64 kbps, the PSTN's own coding), G.729 (8 kbps), or Opus (adaptive, used by many apps).
3. Packetise: every 20 ms of speech goes into one packet with an RTP header (sequence number and timestamp), inside UDP, inside IP. UDP is used because a late voice packet is useless: sending it again would only make things worse.
4. Carry: the packets cross the networks like any others; routers that support quality of service send them first.
5. Play out: the receiver's jitter buffer holds packets for a few tens of milliseconds to even out their varying delay, reorders them by sequence number, covers a lost packet by repeating or smoothing the sound, decodes and plays. RTCP reports loss and delay back to the sender.

Quality depends on three numbers: one-way delay (ITU-T G.114 advises keeping it under about 150 ms for natural conversation), jitter (the variation in delay, smoothed by the buffer) and packet loss (a few percent already sounds broken).

Kinds of VoIP: app to app (WhatsApp, Viber, Messenger, Zoom); IP phones on an office IP-PBX; an analog telephone adapter that lets an ordinary phone use VoIP; gateways that connect VoIP calls to the PSTN, under a softswitch that controls them; and VoLTE, voice over the 4G mobile network, which is VoIP inside the operator's own network.

Advantages: cheap, above all for international calls; one network for voice, video and data; extra features (video, conferencing, voicemail by email, a number that works anywhere). Disadvantages: quality depends on the internet connection and on power; an emergency call cannot easily locate the caller; open to eavesdropping, spam calls and toll fraud unless encrypted and secured (SRTP, TLS).

To remember it: the Viber call a student makes to a parent working in Qatar costs a few megabytes, not an international call rate: the voice was cut into fifty small packets every second and put back together on the other side.

NGN: the next generation network

The problem it solves. A traditional operator ran a separate network for each service: circuit switches for fixed telephones, another core for mobile, another network for data, and often another for television. Each had its own equipment, staff and billing, and a new service had to be built into one of them. An NGN replaces them all with one IP-based packet core that carries voice, video and data together: convergence.

The key ideas:

• Packet-based transport: everything travels as IP packets over one core, often with MPLS underneath for traffic engineering and quality of service (MPLS).
• Service separated from transport: ITU-T Y.2011 splits the network into a transport stratum (moving packets: access and core) and a service stratum (controlling calls and sessions, and providing the applications). A new service can be added without touching the transport, and the transport upgraded without breaking the services.
• Call control in software: a softswitch, or the IP Multimedia Subsystem (IMS) defined by 3GPP, controls calls with SIP, while media gateways connect to the old PSTN.
• QoS-enabled broadband access: DSL, fibre to the home, cable, 4G and 5G radio, all feeding the same core.
• Open interfaces: third parties can build services on the operator's network, and users can reach competing providers.
• Generalised mobility: the same services on any access, fixed or mobile, as the user moves (fixed mobile convergence).

Often drawn as four layers: access (how users connect), transport or core (the IP/MPLS backbone), control (softswitch or IMS) and service or application (voice, video, messaging, third-party applications).

Benefits and challenges: one network to build and run is cheaper, and new services arrive faster; but telephone-grade quality and reliability over shared IP, the security of an open network, and working alongside the old network are all hard.

To remember it: an NGN is a city that replaces its separate pipes for water, gas and drainage with one big service tunnel carrying them all: dig once, maintain once, and add a new service by laying one more line in the same tunnel.

MPLS: forwarding by short labels

Why labels. A normal IP router matches each packet's destination address against a table of perhaps a million routes, looking for the longest matching prefix, at every hop. MPLS does that classification once, at the edge of the network: the packet gets a label, and every router inside simply looks the label up in a small exact-match table, swaps it and forwards. The labels identify virtual paths between distant nodes rather than endpoints, much as a DLCI or a VPI/VCI identifies a circuit.

The label (shim) header, 32 bits placed between the layer 2 header and the IP header:

The parts: a label edge router (LER) at the ingress sorts each packet into a forwarding equivalence class (FEC) (packets to be treated alike, such as one destination prefix with one class of service) and pushes a label; label switching routers (LSRs) in the core swap labels; the egress LER pops the label and forwards by IP. The path a sequence of labels follows is a label switched path (LSP), and it runs one way only. Neighbours agree on labels with LDP, RSVP-TE or BGP.

1. Build the paths: OSPF or IS-IS learns the topology; LDP or RSVP-TE builds the LSPs and the label tables.
2. Push: the ingress LER classifies the packet into an FEC and pushes, say, label 17.
3. Swap: each LSR reads only the label: 17 in becomes 42 out on port 2, and so on.
4. Pop: the egress LER (or the router just before it) removes the label and delivers the plain IP packet.

What it is used for: traffic engineering (steering traffic onto lightly loaded links instead of only the shortest path); MPLS VPNs, which join a company's branches privately across a carrier's shared network and replaced Frame Relay and ATM circuits; quality of service by traffic class; and fast reroute around a failed link within tens of milliseconds. It carries IP, Ethernet, ATM and Frame Relay traffic over T1/E1, ATM, Frame Relay, DSL or Ethernet links: hence multiprotocol.

To remember it: a label is a token number at a busy hospital: the reception reads the whole case once and hands over token 17, and after that every counter reads only the token, not the patient's history.

xDSL: broadband over the telephone line

The last mile. The copper pair from the exchange to a home (the local loop) was built for voice, which needs only about 4 kHz; the wire itself can carry frequencies of a few megahertz over short distances. DSL uses those higher frequencies for data. It covers only the link from the exchange to the home or office, never the links between exchanges: hence a "last mile" technology.

How it works:

• Splitter or microfilter: at the home, it separates the voice band (to the telephone) from the data band (to the DSL modem).
• DSL modem: in the home, puts the data onto the line.
• DSLAM (DSL access multiplexer): at the exchange, ends hundreds of lines and passes their traffic on to the ISP's network.
• DMT modulation (discrete multitone): the band is divided into many 4.3125 kHz subchannels (256 of them in ADSL, up to 1.104 MHz); each carries as many bits as its signal to noise ratio allows, so a noisy part of the band simply carries less.
• Asymmetric speeds: most variants give far more bandwidth downstream (towards the customer) than upstream, which suits browsing and streaming.

Distance is everything: the speed falls as the loop gets longer, because high frequencies fade in copper. ADSL works to about 5 km; VDSL2's top speeds need the customer within a few hundred metres of the equipment.

DSL and ISDN (ISDN) both use the existing copper telephone lines, and both need the customer fairly close to the exchange (the book says usually under 20,000 feet), but DSL is far faster: ISDN's basic rate is 144 kbps (2B+D), while DSL runs to megabits. ADSL traffic between the modem and the exchange was usually carried in ATM cells (ATM). Fibre to the home (FTTH) has since overtaken both.

In Nepal, Nepal Telecom sold ADSL broadband over its landline copper for years; most homes have since moved to fibre to the home from ISPs such as Nepal Telecom, WorldLink and Vianet.

To remember it: the same copper pair that carried grandfather's landline calls carries the internet above them: voice in the bottom 4 kHz, data in the megahertz above, split by a small box at the phone socket.

Chapter 1 in one screen

• Network: autonomous computers and devices joined by links and protocols to exchange data and share resources; parts: nodes, links, protocols, services.
• Uses: business (sharing, reliability, saving money, scalability), home (information, communication, entertainment, e-commerce, education), mobile, society.
• Sizes: PAN (1 to 10 m), LAN (campus), MAN (city), WAN (country); internetwork, the Internet.
• Topologies: bus, star, ring, mesh ($n(n-1)/2$ links), tree, hybrid; physical against logical.
• Client/server: server listens, client requests, server processes and replies; central data, security, backup; single point of failure.
• Peer to peer: every peer client and server; join, search, connect, exchange; BitTorrent, workgroup; cheap, weak security.
• Active network: programmable nodes computing on packets; capsule (integrated) or programmable switch (discrete); NodeOS, EEs, AAs, ANEP.
• Protocol: rules for communication; syntax, semantics, timing; bodies ISO, ITU-T, IEEE, IETF, ANSI, EIA.
• Layering: less complexity, modularity, standards, troubleshooting, reuse; peers, interfaces, virtual communication; architecture = layers + protocols.
• Design issues: addressing, direction, error control, flow control, multiplexing, routing, ordering, segmentation.
• Primitives: LISTEN, CONNECT, RECEIVE, SEND, DISCONNECT; request, indication, response, confirm.
• OSI: physical, data link, network, transport, session, presentation, application (ISO 7498, 1984); bit, frame, packet, segment.
• Which layer: voltage and timing physical; framing, MAC, error control data link; IP address network; socket connection transport; dialog session; encryption presentation.
• TCP/IP: host-to-network, internet (IP), transport (TCP, UDP), application (HTTP, SMTP, DNS, FTP); no session or presentation layer.
• Encapsulation: data, segment, packet, frame (header and FCS trailer), bits; removed in reverse at the receiver.
• OSI against TCP/IP: 7 against 4; model first against protocols first; service, interface, protocol clear against blurred; network layer both against connectionless; transport connection-oriented against both.
• X.25: DTE to DCE interface; physical X.21, LAPB, PLP; LCGN + LCN; Q, D, P(R), M, P(S); call request, incoming call, call accepted, call connected, clear.
• Frame Relay: layers 1 and 2; DLCI (local); PVC and SVC; FECN, BECN, DE, CIR; SVC: SETUP, CALL PROCEEDING, CONNECT, DISCONNECT, RELEASE, RELEASE COMPLETE.
• ATM: 53-byte cells, 5 header + 48 payload; GFC, VPI, VCI, PT, CLP, HEC; AAL1, AAL2, AAL3/4, AAL5; CS and SAR.
• Others: Ethernet IEEE 802.3; VoIP (SIP, RTP over UDP, 80 kbps for G.711); NGN (Y.2001, convergence); MPLS (20-bit label, push, swap, pop); xDSL (ADSL, DSLAM, asymmetric).

The physical layer: moving raw bits as signals PIN 1/27

72 Ka2+6

It moves bits, not meaning. The physical layer does not know whether a bit belongs to an address, a password or a photo; it only makes sure that a 1 sent at one end is read as a 1 at the other. Every frame, packet and segment of the layers above finally travels as these signals.

Four kinds of rule describe a physical interface; standards such as EIA-232 and ITU-T X.21 are written this way:

• Mechanical: the connector's shape, size and number of pins (an RJ-45 jack has eight).
• Electrical: the voltage levels, how long one bit lasts, the longest cable allowed.
• Functional: what each pin or circuit does: transmit data, receive data, clock, ground.
• Procedural: the order of events that brings the link up, uses it and takes it down.

Its functions, the list every answer is built from:

Bit rate is not baud rate. Bit rate counts bits per second; baud rate (signal rate) counts signal elements, or symbols, per second. When each symbol carries $r$ bits, bit rate = baud rate × $r$: a modem sending 2,400 symbols a second at 4 bits a symbol (16 levels) gives 9,600 bps. The bandwidth limits the baud rate, and noise limits how many levels can be told apart (channel capacity).

In the TCP/IP reference model the physical layer is not a layer of its own: the lowest layer, called host-to-network, network access or link layer, covers both the physical and the data link functions. TCP/IP defines no protocol there; it runs over whatever the network below specifies, such as IEEE 802.3 Ethernet over UTP or fiber, IEEE 802.11 Wi-Fi over radio, or DSL over a telephone line. So the functions in the table are the same in both models; only their place differs (Forouzan's five-layer version of TCP/IP does draw a separate physical layer). The models themselves are chapter 1's (TCP/IP, OSI).

Devices that work only at this layer: repeaters, which regenerate a weakened signal; hubs, multiport repeaters that copy every bit to every port; modems, which put bits onto an analog carrier; transceivers; and the cables and connectors themselves (internetworking devices). None of them reads an address.

To remember it: when Bikash sends a file over the hostel LAN, the physical layer of his laptop's network card turns each bit into a voltage pattern on the four copper pairs of the Cat 6 cable, and the switch port in the corridor turns the voltages back into bits. Over the hostel Wi-Fi the same bits leave as 2.4 or 5 GHz radio waves. The file, its packets and its frames are other layers' business: the physical layer only carries ones and zeros.

Delay, latency and throughput: measuring a network PIN 3/27

80 Bh · 78 Bh · 74 Ash1+3+43+3+22+6

Four measures describe how a network performs:

Throughput $={\displaystyle \frac{\text{data delivered}}{\text{time taken}}}$, and it is always at most the bandwidth. It falls below the bandwidth because of headers, retransmissions, collisions, congestion, slow end hosts and the slowest link on the path, the bottleneck. The Ring Road in Kathmandu is the bandwidth; the crawl at rush hour is the throughput. Both board calculations of throughput (2080 Bhadra, 2078 Bhadra) are worked in the Numericals panel; the trap in both is the units: bytes to bits (× 8), minutes to seconds (÷ 60).

The four delays at each node. A packet crossing a router is delayed four times (the book's Figure 2.1):

The signal speed s is about $2\times {10}^8$ m/s in copper and fiber (two thirds of the speed of light) and $3\times {10}^8$ m/s through air and space.

Transmission is not propagation. Transmission delay depends on the size of the packet and the rate of the link, not on the distance; propagation delay depends on the distance and the medium, not on the packet. Picture a bus at the Kathmandu bus park: boarding every passenger through one door is the transmission delay (more passengers or a narrower door, longer), and the drive to Pokhara is the propagation delay (a longer road, a longer drive), whoever is on board.

Queuing delay and traffic intensity. If packets of L bits arrive at an average rate of a packets a second at a link of rate R, the traffic intensity is $La/R$. Near 0 the queue is short; as it approaches 1 the queuing delay grows sharply; above 1 the queue grows without limit and packets are dropped. This is where congestion shows (congestion control).

Latency is the total time from the first bit leaving the source to the last bit arriving: propagation + transmission + queuing + processing, added over every hop. The round-trip time (RTT) is the time there and back, the figure ping reports.

Bandwidth-delay product $=R\times d_{prop}$: the number of bits that fill the link, in flight, before the first one reaches the far end. It says how much a sender may send before an acknowledgment can possibly return, which is why it sizes sliding windows (flow control).

Causes of packet delay, in one list:

• Router processing: header checks, table lookups, encryption on a busy router.
• Congestion: long queues as the traffic intensity nears 1, and loss when a buffer overflows.
• Slow links and big packets: transmission delay L/R at every hop.
• Distance and medium: propagation delay, worst on satellite links.
• Number of hops: store-and-forward means a packet must arrive whole at each router before it is sent on, so every hop adds a transmission delay.
• Retransmissions after errors or loss (error control), and flow and congestion control holding the sender back.
• Slow end systems: a busy server, an overloaded phone.

Jitter hurts voice and video most: a call whose packets leave 20 ms apart but arrive 15, 30 and 18 ms apart sounds broken, so the receiver holds packets in a jitter (playout) buffer to even them out.

Bandwidth and channel capacity: Nyquist and Shannon PIN 1/27

66 Bh4+4

Bandwidth has two meanings. For a signal or a channel it is the width of the range of frequencies passed, in hertz: a telephone voice channel passes about 300 to 3,400 Hz, a bandwidth of about 3.1 kHz (4 kHz with guard bands). For a link it is the bit rate, in bits per second. The two are tied: more hertz allow more bits per second, and Nyquist and Shannon say exactly how many.

Nyquist (1924): the noiseless channel. A channel of bandwidth B can carry at most 2B signal changes a second, and each change between L levels carries ${\log }_{2}L$ bits:

A noiseless 3 kHz channel carries 2 × 3000 × 1 = 6 kbps with 2 levels, and 2 × 3000 × 4 = 24 kbps with 16 levels. More levels raise the rate, but the receiver must still tell them apart, and noise makes levels that are close together impossible to distinguish. That is Shannon's limit.

Shannon (1948): the noisy channel. No number of levels can beat

where SNR is the signal power divided by the noise power, as a plain ratio. In decibels, $SN{R}_{dB}=10{\log }_{10}(S/N)$: 10 dB is a ratio of 10, 20 dB of 100, 30 dB of 1000. Always turn decibels back into a ratio before using Shannon.

• Example, a telephone line: B = 3,000 Hz and SNR = 30 dB (1,000): $C=3000{\log }_{2}1001$ ≈ 29.9 kbps. That is why dial-up modems on analog lines stopped near 33.6 kbps.
• Using both together: Shannon gives the ceiling, Nyquist the levels needed. For B = 2 MHz and SNR = 255, Shannon gives $2\times {10}^6{\log }_{2}256$ = 16 Mbps. Choosing a safer 12 Mbps, Nyquist gives $12\times {10}^6=2\times 2\times {10}^6{\log }_{2}L$, so ${\log }_{2}L=3$ and L = 8 levels.

Why the received signal is worse than the one sent (transmission impairments):

Loss in decibels is $10{\log }_{10}(P_{out}/P_{in})$: a signal that falls to half its power has lost about 3 dB.

The board question (2066 Bhadra) gives the signal as 2 mW and the noise as 200 µW: put both in milliwatts first (200 µW = 0.2 mW), so SNR = 10, which is 10 dB; then $C=300\times {10}^6{\log }_{2}11$ ≈ 1.04 Gbps. It is worked in full in the Numericals panel.

To remember: bandwidth is the width of the road, the number of levels is how many lanes the drivers can keep apart, and noise is the dust that makes lane markings unreadable; Shannon is the road authority's limit that no lane painting can beat.

Optical fiber: light in glass PIN 1/27

82 Bh4+2

Structure, from the centre out: the core of glass (8 to 10 µm across in single mode fiber, 50 or 62.5 µm in multimode); the cladding, glass of a lower refractive index, 125 µm across; a plastic buffer coating, 250 µm; strength members of aramid yarn (Kevlar); and the outer jacket. The cross section is drawn on the guided media card (guided media).

How the light stays in: total internal reflection. Light passing from a denser medium (the core, index $n_1$) into a rarer one (the cladding, $n_2$, smaller than $n_1$) bends away from the normal. Beyond the critical angle it does not cross at all: it is reflected back into the core, and does so again at every bounce along the fiber.