Study the network as a system, not a list of definitions.

A computer network connects autonomous devices so they can exchange information and share services. Its value comes from communication, access, coordination, and resource sharing rather than from the cables or radio links alone.

Network hardware is commonly classified by transmission technology and geographical scale. A broadcast link is shared by many devices; a point-to-point link directly connects a pair of devices. Delivery may be unicast, multicast, broadcast, or anycast.

Internet as a network of networks

The Internet does not use one single owner or one physical technology. Internet service providers, enterprise networks, access networks, data centres, and backbone networks exchange packets using the TCP/IP protocol suite. Routers connect networks and choose packet paths.

Network software is organized into layers. Each layer performs a related set of functions, offers services to the layer above it, and uses services of the layer below it. A protocol specifies the rules and formats used by peer entities in the same layer; an interface defines how adjacent layers interact.

Why layering is used

• Reduces design complexity by separating concerns.
• Allows a layer implementation to change without redesigning every other layer, provided its service interface remains stable.
• Promotes interoperability and standardization.
• Makes testing, troubleshooting, and protocol evolution easier.

Encapsulation and virtual communication

When data moves downward, each layer may add control information called a header (and sometimes a trailer), creating its protocol data unit. At the receiver, the process is reversed. Peer layers appear to communicate directly, but actual bits travel through lower layers and the physical medium.

Although layers have different purposes, several design questions recur throughout a network architecture.

A service describes the behaviour visible to the service user. It may be connection-oriented, where state is established before transfer, or connectionless, where each message is treated independently.

A service is accessed through operations called service primitives. A simple connection-oriented interface may use LISTEN, CONNECT, RECEIVE, SEND, and DISCONNECT.

The ISO Open Systems Interconnection model organizes networking into seven conceptual layers. It is primarily a reference and teaching model: each layer groups well-defined functions and communicates through interfaces.

The TCP/IP model grew from operational internetworking and the ARPANET. In the four-layer version used by the course, the layers are Link, Internet, Transport, and Application.

IP provides an unreliable best-effort datagram service. Reliability, ordering, and congestion/flow functions may be added at the transport layer by TCP. UDP provides a lighter connectionless transport service.

Critique of OSI

• Bad timing: the protocol suite competed with already-growing TCP/IP deployment.
• Bad technology: some functions and layer boundaries were considered complex or awkward.
• Bad implementations: early implementations were large and slow.
• Bad politics: TCP/IP was associated with open academic/Internet growth while OSI was perceived as committee-driven.

Critique of TCP/IP

• The model does not cleanly distinguish services, interfaces, and protocols.
• The host-to-network/link layer is underspecified.
• It is less general as a reference model because it closely follows one protocol suite.
• Some layer functions are not separated as cleanly as in OSI.

Example networks show how architectural ideas become real systems. ARPANET demonstrated packet switching and internetworking; NSFNET expanded academic backbone connectivity; the modern Internet is a hierarchy of access networks, regional providers, backbone providers, Internet exchange points, and data centres.

Why standards matter

Standards let equipment and software from different vendors interoperate. De facto standards become accepted through widespread use; de jure standards are formally approved by recognized organizations.

The data link layer converts the raw service of the physical layer into a link service for the network layer. It accepts packets, places them in frames, coordinates transmission over one link, and hands received packets upward.

The physical layer supplies a continuous bit stream. Framing divides that stream into recognizable units so headers, payload, and trailer information can be interpreted correctly and the receiver can resynchronize after disturbances.

1. Byte count

A header field gives the frame length. The receiver counts that many bytes to find the next frame. Its weakness is loss of synchronization when the count field is corrupted.

2. Flag bytes with byte stuffing

Special FLAG bytes mark the beginning and end. If FLAG or ESC occurs inside the data, the sender inserts ESC before it. The receiver removes the inserted ESC during destuffing.

3. Flag bits with bit stuffing

A bit pattern, commonly 01111110, acts as a flag. After five consecutive 1s in payload data, the sender inserts a 0. The receiver removes that 0. Thus the flag pattern cannot appear unintentionally inside data.

Flow control prevents a fast sender from delivering frames faster than a receiver can accept, buffer, and process them. It concerns receiver capacity; it must not be confused with congestion control, which concerns overload inside the network.

Utopian simplex protocol

Assumes data travels in one direction, the channel never damages or loses frames, the receiver is always ready, and its buffer is unlimited. The sender repeatedly obtains a packet, places it in a frame, and transmits. It is a baseline rather than a practical protocol.

Simplex stop-and-wait for an error-free channel

Removes the assumption of an infinitely fast receiver. The sender transmits one frame and waits for an acknowledgement before taking the next packet. This provides feedback-based flow control but wastes capacity on long-delay links.

Simplex stop-and-wait for a noisy channel

Adds sequence numbers, acknowledgements, a timer, and retransmission. If a data frame or ACK is lost, the sender eventually times out and sends again. A one-bit sequence number is sufficient because only one frame can be outstanding; it lets the receiver detect duplicates.

A sliding-window protocol allows sequence numbers within a moving range. The sender window contains frames that may be transmitted or are awaiting acknowledgement; the receiver window contains sequence numbers it is prepared to accept.

Piggybacking

On a full-duplex link, an acknowledgement can be carried in the header of a reverse-direction data frame rather than sent separately. A short timer prevents an ACK from being delayed indefinitely when reverse traffic is absent.

One-bit sliding window

With sequence numbers 0 and 1 and maximum window size 1, both sides alternate expected numbers. This is essentially bidirectional stop-and-wait with piggybacking. Duplicate frames are recognized by the sequence bit.

Go-Back-N pipelines frames. The sender may have up to W unacknowledged frames, while the receiver window is 1. The receiver accepts only the next expected frame and discards later out-of-order frames. ACKs are cumulative.

Window-size rule

For an m-bit sequence number, Go-Back-N normally uses a sender window no larger than 2^m − 1. Reserving one sequence number prevents an old frame from being confused with a new frame after wraparound.

Selective Repeat retransmits only frames that are lost or damaged. The receiver accepts and buffers valid out-of-order frames, acknowledges them individually, and delivers them upward after the missing gap is filled.

Packet over SONET

SONET is a synchronous optical transport system used in backbone networks. IP packets can be carried through PPP framing over SONET. The encapsulation path is conceptually IP packet → PPP frame → SONET payload.

ADSL

Asymmetric Digital Subscriber Line uses existing telephone copper to provide a higher downstream rate than upstream rate. A typical path is PC → Ethernet → DSL modem → local loop → DSLAM at the provider → ISP network. ADSL systems may carry PPP over ATM/AAL5 in the architecture shown in the lecture.

When many stations share one broadcast channel, the MAC sublayer decides who may transmit. If two stations transmit at an interfering time, a collision can destroy useful work.

Assumptions used to analyse dynamic protocols

• Independent stations generate frames.
• One shared channel is available.
• Collisions are observable or inferable.
• Time may be continuous or slotted.
• Carrier sensing may be present or absent.

ALOHA is a random-access protocol: a station transmits when it has a frame and retransmits after a random delay if a collision is inferred.

Pure ALOHA

Transmission may begin at any time. A frame of duration T is vulnerable to another frame beginning during the interval from T before its start to T after its start, so the vulnerable period is 2T.

Slotted ALOHA

Time is divided into slots of one frame duration and transmission begins only at a slot boundary. Synchronization reduces the vulnerable period to T, improving maximum throughput.

Carrier Sense Multiple Access listens before sending. Collision Detection adds the ability to monitor the medium while transmitting, abort when a collision is detected, send a jam signal, and retry after binary exponential backoff.

Why the minimum frame length exists

The sender must still be transmitting when a worst-case collision propagates back from the farthest station. Therefore frame transmission time must be at least the round-trip propagation time: L/R ≥ 2Tprop, or Lmin = 2RTprop.

In the basic bit-map protocol, a contention period contains one reservation bit per station. Station i sets bit i if it has a frame. After the reservation phase, ready stations transmit in numerical order without collisions.

Performance intuition

• At high load, reservation overhead is shared across many transmitted frames.
• At low load, scanning N reservation bits for only a few frames creates significant delay.
• The protocol is fair by station order but higher-numbered stations may wait longer.

Stations are logically arranged in a ring. A special control frame called a token circulates. Only the station holding the token may transmit, so simultaneous transmissions do not collide.

Operation

1. A station with no frame passes the token.
2. A station with data captures the token and sends within the token-holding limit.
3. After completing transmission, it releases a new token.
4. Monitoring procedures recover from a lost token or duplicate tokens.

Ready stations transmit their binary addresses one bit at a time, usually from the most significant bit. A dominant bit value overwrites the other value on the shared channel. A station that observes a higher-priority bit withdraws; the highest address remains and transmits.

Characteristics

• Collision resolution requires log₂N address-bit times.
• Priority is deterministic, so fixed addresses can starve low-priority stations.
• Rotating or virtual priorities can improve fairness.

Contention protocols perform well at low load because a station can transmit quickly. Collision-free protocols perform well at high load because they schedule access. Limited-contention protocols adapt between these extremes.

Adaptive tree walk

Stations correspond to leaves of a binary tree. A contention slot tests a node representing a group of stations. No transmission means the group is empty; one transmission succeeds; a collision causes the algorithm to test the node's child groups. Group size can be chosen according to load.

Classic Ethernet used a shared coaxial cable or hubs, Manchester encoding at 10 Mbit/s, and 1-persistent CSMA/CD. Every station in the collision domain saw the shared signal.

The frame from destination address through FCS is at least 64 bytes and at most 1518 bytes without VLAN tagging. The minimum keeps a CSMA/CD sender active long enough to detect a worst-case collision.

A hub repeats signals into one shared collision domain. A switch learns MAC addresses and forwards frames only toward the appropriate port, giving each full-duplex link its own collision domain. Full-duplex switched Ethernet does not need CSMA/CD.