The Link Layer, LANs and Wireless Access

A datagram may cross Wi-Fi, Ethernet and several router links on its way to a server. The link layer handles each local crossing. Its packet, a frame, wraps the datagram in a header and usually a trailer.

Follow the next hop

A node is a device running a link-layer protocol: a host, router, switch or access point. A link connects adjacent nodes. The network layer chooses where the datagram should go next; the link layer delivers it across the selected link.

Imagine a laptop sending through an access point and a campus router. The laptop first sends an 802.11 frame to the AP. The AP bridges the traffic onto Ethernet. The campus router removes the incoming frame, examines the IP destination and builds a frame suitable for its outgoing link. A switch within the Ethernet LAN forwards a frame without replacing its source and destination MAC addresses. A router creates a new link-layer envelope for the next network.

Four services, different choices

Protocols choose the services their links need. Ethernet checks a frame's CRC and discards a damaged frame without acknowledging it. Wi-Fi acknowledges correctly received unicast data frames; a missing ACK can trigger a local retry. Wireless errors are common enough that recovering on the wireless hop can save the delay of an end-to-end TCP retransmission. A successful Wi-Fi exchange still does not promise that a datagram will survive all later links.

Shared and dedicated channels

A point-to-point link joins two endpoints, such as a host and one switch port. A shared or broadcast medium supports several transmitters, as on a Wi-Fi channel or a cable internet upstream. They need rules for sharing airtime. A shared channel does not guarantee that all stations can hear each other: obstacles and distance can hide wireless transmitters.

Full duplex lets both endpoints send at the same time. Half duplex lets only one send at a time. A modern switched Ethernet host link is normally full duplex; Wi-Fi stations share radio airtime.

Where the work happens

The network interface controller (NIC) performs framing, medium access and error checks in hardware at line speed. Operating-system software supplies datagrams and addressing information, handles interrupts and passes received datagrams upward. To send, the NIC builds the frame, computes check bits, follows the access rules and transmits. To receive, it checks the frame and extracts the datagram if the frame is acceptable.

Keep the two address scopes separate. Link-layer addresses identify the sender and receiver for a local delivery. IP source and destination addresses normally identify the endpoints across the routed path; NAT can rewrite them.

Error checks add information that the receiver can test. A failed check proves damage; a passing check means that the scheme found no error. Different checks miss different patterns.

Count ones with parity

With even parity, choose a bit that makes the total count of ones even. The data 1101011 has five ones, so append 1: the transmitted sequence is 11010111. If one bit flips, the count becomes odd and the receiver detects it. An odd number of flips always changes parity. An even number preserves it, including two flips that change a zero to one and a one to zero.

Two-dimensional parity checks both rows and columns of a block. For rows 1011, 0110 and 1100, the even row-parity bits are 1, 0 and 0. Column-parity bits are 0, 0, 0 and 1; a corner bit completes parity for the check bits. If row 2 becomes 0100, row 2 and column 3 fail. Their intersection locates the changed bit, which the receiver flips back.

The repair rule is for a single-bit error. Two-dimensional parity detects all one-, two- and three-bit error patterns, but multiple failures need not identify a unique repair. Four flipped bits at the corners of a rectangle change two bits in each affected row and column. Every parity check then passes.

Repair without a round trip

Forward error correction (FEC) sends enough redundancy for the receiver to recover some damaged data without asking for another copy. This is useful when a retransmission would arrive too late, as with real-time traffic or a long satellite round trip. Extra check bits consume capacity even when no error occurs. Modern FEC codes offer stronger protection than the simple row-and-column scheme.

Add 16-bit words

The Internet checksum uses ones' complement addition. Add the 16-bit words, then wrap any overflow carry around into the low-order end. Invert every bit of the final 16-bit sum to form the checksum.

For 0x8F3A, 0x6C21 and 0x1E07:

1. Add the first two words: 0x8F3A + 0x6C21 = 0xFB5B.
2. Add the third: 0xFB5B + 0x1E07 = 0x11962.
3. Fold the carry: 0x1962 + 0x0001 = 0x1963.
4. Invert the bits: the checksum is 0xE69C.

The receiver includes the checksum in its sum. Here, 0x1963 + 0xE69C = 0xFFFF, so the check passes. Dropping the overflow would produce the wrong checksum; wrapping is part of the arithmetic.

The checksum is cheap in software and is used by IPv4, TCP and UDP. Its coverage differs: IPv4 checks its header; TCP and UDP cover their transport data and a pseudo-header. Swapping complete 16-bit words leaves the sum unchanged, so the checksum cannot detect that reordering. Ethernet and Wi-Fi use a stronger CRC computed efficiently in adapter hardware.

A cyclic redundancy check (CRC) treats bits as coefficients in a polynomial calculation. You can compute it directly with bit patterns: shifts and exclusive OR replace ordinary multiplication and subtraction.

Set the check width

Sender and receiver agree on a generator, G. A generator with r + 1 bits produces r check bits. For G = 1011, r is 3. The sender chooses those three bits so that data followed by the check bits divides exactly by G in modulo-2 arithmetic.

Modulo-2 addition and subtraction are both XOR. Equal bits give 0; unequal bits give 1. Thus 1011 XOR 0101 = 1110. There are no carries or borrows. Appending r zeros shifts a data pattern D left by r places, equivalent to multiplying by $2^r$.

Work the division

Use D = 11010011 and G = 1011.

1. Append three zeros: 11010011000.
2. Starting at the leftmost 1, align the leftmost generator bit with it and XOR.
3. Move right to the next remaining 1 and repeat. Skip leading zeros.
4. Stop after processing the data positions; the last three bits are the remainder R.

The successive full-width patterns are:

11010011000
01100011000
00111011000
00010111000
00000001000
00000000011

The remainder is 011. Keep its leading zero: the check field always has three bits. Replace the appended zeros with this remainder and transmit 11010011011.

Check what arrived

The receiver divides the complete received sequence by the same generator. It does not append another set of zeros. Dividing 11010011011 gives remainder 000, so no error is detected. If the third bit changes, the received sequence becomes 11110011011; its remainder is 010, so it is rejected.

This works because the valid transmitted pattern is a multiple of G. An error changes that pattern. The receiver detects the change unless the error pattern is also a multiple of G.

Know the protection

A suitable degree-r generator detects every burst error whose span is at most r bits. For longer random bursts, the chance of an undetected error is about $2^{-r}$. Generators containing the factor $x+1$ also detect every odd number of flips. These are different guarantees: burst length is the distance from the first changed bit to the last, including unchanged bits between them.

Ethernet and Wi-Fi use a 32-bit CRC, carried as a four-byte frame check sequence (FCS). Its random long-burst miss probability is about one in 4.3 billion. CRC-32 detects up to three flipped bits in an Ethernet-size frame. Hardware shift registers and XOR gates can compute it as bits pass through the adapter.

A CRC detects accidental damage; it does not authenticate a sender. Someone deliberately altering a frame can calculate a new CRC. It also supplies no automatic repair: Ethernet drops a bad frame, while Wi-Fi's missing ACK can cause a retry.

Several interfaces sharing one channel need a rule for who sends. If transmissions overlap at a receiver and cannot be decoded, a collision wastes channel time. A medium access protocol coordinates the competing senders.

What would ideal sharing achieve?

For a channel of rate R, one active sender should get all of R. With M active senders, each should average R/M. Coordination should be simple, inexpensive and decentralized, without a vulnerable controller or synchronized clocks. Real protocols trade among these aims.

Three families make different choices: channel partitioning reserves shares, random access allows attempts and resolves collisions, and taking turns grants temporary permission to send.

Reserve time or frequencies

TDMA assigns one slot per node in each repeating time frame. Each transmission uses the full channel rate during its slot, but a node's long-run share is R/N for N assigned nodes. Four nodes sharing 1 Gb/s each average 250 Mb/s. If three have no data, their fixed slots remain empty; the fourth still averages 250 Mb/s.

FDMA assigns a separate frequency band to each node. The node can transmit continuously in its band, but an equal division again gives R/N. Silence wastes a reserved band just as it wastes a TDMA slot.

Fixed partitioning prevents contention among assigned users, but bursty traffic often leaves shares idle while another sender has a queue.

Separate overlapping signals with codes

CDMA represents a data bit as +1 for bit 1 or −1 for bit 0. Multiply that value by a sender-specific sequence of chips, each +1 or −1. Simultaneous chip sequences add in the channel.

Take codes $c_1=(1,1,1,1)$ and $c_2=(1,-1,1,-1)$. Their inner product is zero, so they are orthogonal. If sender 1 transmits +1 and sender 2 transmits −1, their chip sequences are (1, 1, 1, 1) and (−1, 1, −1, 1). The received sum is (0, 2, 0, 2).

To decode sender 1, multiply received chips by its code, add, and divide by four: $(0+2+0+2)\mathrm{/}4=1$. Sender 2's calculation is $(0-2+0-2)\mathrm{/}4=-1$. Each code removes the other sender's contribution in this ideal model.

Actual systems need careful power control so that a nearby strong signal does not overwhelm a distant weak one. Orthogonality is a property of code pairs, not a promise that arbitrary radio signals separate perfectly. CDMA underpinned 3G access; LTE and 5G use OFDM-based access, with LTE's uplink using the related SC-FDMA scheme.

Random access lets a node transmit without first reserving a share. A collision is followed by an independently chosen random delay. Randomness helps senders stop repeating the same collision together.

Add slot boundaries

In slotted ALOHA, equal-length frames fit exactly into slots of L/R seconds. Nodes synchronize their clocks and begin only at slot boundaries. A new frame uses the next slot. After a collision, a waiting node attempts each later slot with probability p.

A slot has three possible outcomes: no node transmits, exactly one transmits successfully, or several collide. Making p very small reduces collisions but creates empty slots. Making it too large fills slots with competing attempts.

Derive success probability

Suppose N nodes always have frames and attempt independently with probability p. For one named node to succeed, it must transmit and all N − 1 others must remain silent. Its success probability is $p(1-p{)}^{N-1}$. Any one of the N nodes can be that sole sender, giving total efficiency

With N = 10 and p = 0.1, the value is $10(0.1)(0.9{)}^9\approx 0.387$. A 100 Mb/s channel therefore carries about 38.7 Mb/s of successful frames in aggregate. This is not 38.7 Mb/s for each node. Symmetric saturated nodes share that successful capacity.

The best attempt probability is $p=1\mathrm{/}N$. As N grows, the maximum useful fraction approaches $1\mathrm{/}e\approx 0.368$. The missing fraction includes both collisions and empty slots. Random access has overhead even when all senders are ready.

Remove synchronization

Pure ALOHA sends whenever a frame is ready, without slots. Suppose a frame takes T seconds. Another transmission beginning just before it can overlap its start; one beginning before it finishes can overlap its end. The vulnerable interval spans 2T.

Slots remove partial overlaps by forcing every start to a boundary. In the standard saturated probability model, a pure-ALOHA sender succeeds with probability $p(1-p{)}^{2(N-1)}$, and the large-population maximum aggregate efficiency is $1\mathrm{/}(2e)\approx 0.184$. Pure ALOHA is simpler to coordinate but has about half the best throughput of slotted ALOHA.

A single active sender can use every slot in slotted ALOHA. The 37% limit describes many competing saturated nodes, not a physical ceiling on every use of the protocol. ALOHA began as an island radio network in Hawaii; its contention-and-retry idea also appears in cable and cellular access requests.

Carrier sense multiple access (CSMA) makes a node listen before transmitting. Collision detection (CD) makes a transmitting node stop when it detects a collision. Listening improves efficiency but cannot eliminate the delay before a distant signal arrives.

Why an idle channel can still collide

If B begins sending at one end of a cable, D at the other end may still hear silence. D can begin during that propagation interval, producing a collision. With CSMA/CD, adapters monitor the medium while sending. On detecting a collision, an adapter stops the frame, sends a jam signal and schedules a retry.

Choose a random retry delay

After collision n for the same frame, binary exponential backoff chooses an integer K uniformly from 0 through $2^m-1$, with $m=\min (n,10)$. The delay is K times 512 bit times. The range doubles after successive early collisions, spreading competing retries over more possibilities. The adapter abandons a frame after 16 unsuccessful attempts.

At 100 Mb/s, one bit takes 10 ns. A 512-bit slot therefore lasts 5.12 μs. After the third collision, K can be 0 through 7, so the longest wait is $7(5.12)=35.84$ μs. At 10 Mb/s, the same 512 bit times last 51.2 μs. A bit-time count is not a fixed number of seconds across link speeds.

Check the network's timing

Collision detection works only while the sender is still transmitting. A minimum frame must outlast the worst-case trip to a distant collision and back. Classic Ethernet's 64-byte minimum supplies 512 bit times.

For a 2 km span and signal speed $2\times {10}^8$ m/s, the round-trip propagation time is $4000\mathrm{/}(2\times {10}^8)=20$ μs. That exceeds a 100 Mb/s slot's 5.12 μs, even before device delays. A minimum frame could finish too early. The propagation-only distance bound is about 512 m; real shared-network limits must also allow for hubs and adapters.

Estimate efficiency

A common approximation is

where propagation delay is one-way across the network and transmission delay is the time to send a maximum-size frame. For 200 m at the same signal speed, propagation takes 1 μs. A 1,518-byte frame at 100 Mb/s takes $1518(8)\mathrm{/}{10}^8=121.44$ μs. Substitution gives about 0.96, or 96% efficiency. Smaller propagation delay and longer frames reduce the relative cost of contention.

Modern switched, full-duplex Ethernet has no shared-medium collisions and does not run CSMA/CD. Ethernet at 10 Gb/s and above is full duplex only. Wi-Fi also senses the carrier, but uses collision avoidance: a transmitting radio cannot reliably detect another station's much weaker signal, and hidden stations may be out of hearing range.

Taking-turns protocols try to give a lone sender nearly the whole channel while dividing capacity fairly when many senders are active. They avoid repeated collisions by controlling permission to transmit.

Ask each sender or pass permission

In polling, a controller invites nodes to send in sequence, often with a limit on frames per turn. The limit prevents one busy node from keeping everyone else waiting. Poll messages consume airtime, and a ready node may wait while the controller checks idle nodes. The controller is also a single point of failure. Classic Bluetooth uses centrally coordinated access in a piconet.

In token passing, a small control frame circulates in a fixed order. Holding the token permits a node to send a bounded amount. A node with no data passes it onward promptly. No central poller is needed, but the protocol needs procedures for a lost token or a break in the logical cycle. Token Ring and FDDI used this approach.

On a ten-node channel with only node 7 active, fixed TDMA grants it one tenth of the time. Slotted ALOHA can let it use every slot. Polling approaches the full rate but spends some time polling the nine silent nodes.

Follow cable upstream access

DOCSIS connects cable modems to a cable modem termination system (CMTS). Frequency resources separate channels. On an ordinary downstream channel, the CMTS is the single sender. On an upstream channel, many modems compete to reach that same CMTS.

The upstream is divided into minislots. A downstream MAP message announces which upstream opportunities belong to which modems. The CMTS can give one modem consecutive minislots for a data burst, then give the next opportunity to another modem.

A modem with queued data can send a short bandwidth request in a contention opportunity. Requests can collide. The modem monitors later MAPs for its request's outcome. Once the acknowledgment time has passed its request and no grant or pending grant covers it, the modem retries using a truncated binary exponential backoff. Merely lacking an immediate data grant does not prove a collision: a recognized request may still be waiting for scheduling.

DOCSIS therefore combines three ideas: frequency division organizes channels, random access obtains requests, and scheduled grants carry upstream data. DOCSIS 3.1 and 4.0 add OFDM downstream and OFDMA upstream, dividing channels into narrow subcarriers while retaining request-and-grant access. Calling cable access only TDMA misses the contention that obtains a transmission opportunity.

An IP address helps select a network destination. A MAC address identifies an interface for local frame delivery. A laptop with Ethernet and Wi-Fi has a separate MAC address for each interface.

Read a 48-bit address

A common MAC address has six bytes, written as six hexadecimal pairs: 5C-F9-38-A2-41-07. The 48-bit space contains about 2.8 × 10¹⁴ values. In the familiar manufacturer assignment model, an IEEE-issued organizationally unique identifier (OUI) supplies the first 24 bits and the manufacturer assigns the remaining 24.

MAC addresses are flat: no part specifies the interface's current IP subnet. An interface can retain its MAC address after moving. Software can also replace a factory address; Wi-Fi devices commonly use randomized locally administered addresses for privacy. Addresses must remain unique within the local LAN.

The two low-order bits of the first byte have special meanings. The least significant bit is 0 for an individual address and 1 for a group address. The next bit is 0 for universal administration and 1 for local administration. 5C ends in binary 00, so it identifies a universally administered unicast address. DA ends in 10, so it identifies a locally administered unicast address. FF-FF-FF-FF-FF-FF is the broadcast destination.

Resolve an IPv4 neighbor

Address Resolution Protocol (ARP) maps an IPv4 address to a MAC address on the local subnet. An ARP cache stores IP–MAC mappings with expiration information. ARP messages travel directly in link-layer frames, identified by EtherType 0x0806.

Suppose A at 192.168.1.10/24 needs B at 192.168.1.20 and has no cached mapping:

1. A broadcasts an ARP request asking who owns B's IPv4 address. It includes A's own IP and MAC addresses.
2. All local interfaces receive the request; B recognizes the target address.
3. B sends its reply to A's MAC address and can learn A's mapping.
4. A caches B's mapping and sends the data frame to B.

The destination adapter accepts a frame addressed to itself, the broadcast address, or a multicast group it receives. It normally discards frames addressed elsewhere before they reach IP.

Send beyond the subnet

For destination 203.0.113.5, A first recognizes that the address is outside its /24 subnet. It selects gateway 192.168.1.1, resolves that gateway's MAC address and sends the Ethernet frame there. The enclosed IP datagram remains addressed to 203.0.113.5. Each router makes its next-hop decision and builds the outgoing frame.

ARP cannot discover a remote server's MAC address through routers. Routers stop the local broadcast, and the remote MAC would not identify the next receiver on A's cable anyway.

ARP does not authenticate ownership claims, so forged replies can redirect local traffic. IPv6 uses ICMPv6 Neighbor Discovery, with neighbor solicitations and advertisements, instead of ARP.

Ethernet spans many physical media and speeds, while retaining a recognizable frame structure and 48-bit addressing. Shared coaxial Ethernet and hub-based Ethernet used one collision domain. A switch gives each host a separate link, normally operated full duplex.

Read the fields

The synchronization sequence comprises seven alternating-bit bytes and a start delimiter. It precedes the part counted in the usual 64- to 1,518-byte frame length. EtherTypes include 0x0800 for IPv4, 0x86DD for IPv6 and 0x0806 for ARP. The receiver uses this field to hand the payload to the correct protocol.

Separate MTU from frame size

The usual Ethernet maximum transmission unit (MTU) is 1,500 bytes: the largest network-layer datagram carried in the data field. It excludes Ethernet's header and CRC. A larger datagram needs a compatible larger MTU or network-layer handling such as smaller packets; IPv4 can fragment where permitted, while IPv6 routers do not fragment forwarded packets.

A 20-byte datagram needs 26 padding bytes to reach the 46-byte minimum data field. Adding 14 header bytes and four CRC bytes gives the 64-byte minimum frame. The receiver uses the IP packet length to distinguish the datagram from padding. At the other extreme, a 1,500-byte datagram gives a 1,518-byte untagged frame.

Include occupied link time

Every frame also uses the eight-byte preamble/start sequence and a 96-bit idle gap, equivalent to 12 byte times. A maximum-size transmission therefore occupies 1,538 byte times. The data-field fraction is $1500\mathrm{/}1538\approx 0.975$, about 97.5%.

A minimum frame occupies $64+8+12=84$ byte times. Its 46-byte data field is about 55% of that time. If only 20 bytes are actual IP data, useful IP efficiency is lower still: padding is not application information. Always name which bytes count as useful when reporting efficiency.

Understand the service

Ethernet is connectionless: no Ethernet handshake precedes a frame. It is also unreliable in the protocol sense: a bad CRC causes a drop without an Ethernet ACK or retry. TCP or an application may recover the lost data later.

IEEE 802.3 includes 10 Mb/s through 800 Gb/s technologies over copper and fiber. Names such as 100BASE-TX and 10GBASE-T encode speed and physical signaling details. Some data centers use jumbo frames with roughly 9,000-byte payloads. Support must agree along the relevant path; a faster link alone does not enlarge the normal MTU.

A switch connects Ethernet segments and forwards frames by destination MAC address. Hosts address each other or a router, not the switch. This transparency lets a switch be inserted without changing the hosts' addressing.

Make two separate decisions

For every arriving frame, first learn the source: record its MAC address, incoming port and current time. Then look up the destination to decide where the frame should go.

Learning from the destination would be circular: the switch has not observed where that destination actually sends. Entries age out when they are not refreshed, commonly after about 300 seconds, so old locations do not remain forever.

Trace an initially empty table

Connect A to port 1, B to port 2 and C to port 3.

1. A sends to B. The switch records A → 1. B is unknown, so the frame leaves ports 2 and 3.
2. B replies to A. The switch records B → 2 and sends only to port 1.
3. C sends to B. The switch records C → 3 and sends only to port 2.

A broadcast is flooded even when other addresses are known. An unknown unicast is also flooded, but it remains addressed to one interface; other adapters discard it. A silent destination stays unknown until a frame with its source address reaches the switch.

If A and D share a hub connected to port 1 and D is already learned there, A's frame for D arrives on port 1. The switch filters it because D already lies on the incoming segment. D can receive the hub's local copy without the switch sending another copy.

Avoid endless flooding

Ethernet frames have no hop-count or TTL field. A loop can send broadcasts around repeatedly, multiplying copies into a broadcast storm. Spanning Tree Protocol (STP) creates a loop-free active tree by blocking redundant links. A blocked path can become active after a failure. In an STP-based layer-2 topology, redundant links provide resilience but are not all used simultaneously for ordinary forwarding.

Compare the boundaries

Switches separate collision domains and can connect ports of different speeds. They may buffer incoming frames and queue them at busy output ports. They do not remove a bottleneck when several inputs compete for one slower output.

Routers forward by IP, separate subnets and stop link-layer broadcasts. Their hierarchical addressing and routing protocols support larger networks; IP TTL limits looping datagrams. Switched LANs instead need a loop-free forwarding topology and control of broadcast growth.

Self-learning also exposes a weakness: fake source addresses can exhaust a switch's table. More legitimate destinations then become unknown and may be flooded. VLAN boundaries still limit the flood's scope; table learning is not an authentication mechanism.

One large switched LAN spreads broadcasts to many hosts and ties logical groups to physical wiring. A virtual LAN (VLAN) divides the same switching infrastructure into separate layer-2 networks.

Assign logical membership

With port-based VLANs, each access port belongs to a chosen group. Ports 2–8 might belong to engineering VLAN 10 and ports 9–15 to finance VLAN 20. A broadcast entering VLAN 10 reaches the other VLAN 10 ports, not the finance ports. Moving a user between groups can be a software change to port membership.

A VLAN is a broadcast domain. It can span several physical switches, and one switch can support several VLANs. Hosts on ordinary access ports send and receive ordinary untagged frames; they need not know how the switches divide the LAN.

Route between groups

Different VLANs cannot exchange frames directly. An engineering host at 10.1.10.25 sending to finance host 10.1.20.40 recognizes a remote subnet and sends to its gateway's MAC address. The router removes the VLAN 10 frame and forwards the IP datagram into VLAN 20 in a new frame. A layer-3 switch can perform this routing internally.

Separation at layer 2 does not itself forbid routed communication. Filtering rules on the routed path decide which inter-VLAN traffic is allowed. Assigning two ports to different VLANs is not a replacement for that policy.

Carry membership over a trunk

A trunk carries several VLANs between switches, limited to those allowed on it. In the tagged case, IEEE 802.1Q inserts four bytes after the source MAC address:

• A two-byte tag protocol identifier, commonly 0x8100.
• Three priority bits.
• One drop-eligible bit.
• A 12-bit VLAN identifier, or VID.

The receiving switch uses the tag to keep the frame in its correct VLAN. Delivery to an ordinary untagged access port removes the tag. Trunks can also support an untagged native VLAN; tagging rules must be consistent between endpoints.

The VID has $2^{12}=4096$ bit patterns. Values 0 and 4095 are reserved, leaving 4094 ordinary VLAN IDs. A tag expands a maximum untagged 1,518-byte frame to 1,522 bytes. Changing the bytes covered by the FCS requires a new CRC.

Extend isolation across a routed fabric

Cloud networks may need more than 4094 isolated groups. VXLAN encapsulates an Ethernet frame in UDP and IP, using a 24-bit virtual network identifier. This supplies about 16.7 million identifiers while the underlying routers forward the outer IP packet.

The inner frame expresses the virtual LAN's local delivery; the outer packet crosses the physical routed fabric. This allows virtual networks to span servers without making the entire data center one giant physical broadcast domain. VLANs can also be assigned by other criteria, such as source MAC address or protocol, although port membership is the simplest model.

A data center serves external clients and also moves large amounts of traffic between its own servers. Storage replication, search fan-out and distributed machine learning can all stress links between racks.

Build the fabric

Servers connect to a top-of-rack (ToR) switch, often one for 20–40 servers. In a classic hierarchy, ToR switches connect upward through tier-2 and tier-1 switches. Border routers connect the fabric to the internet. A load balancer receives traffic for a service and distributes requests among servers; clients can use one public service address while the internal server set changes.

The critical question is how much traffic can leave a rack. Having fast server links does not guarantee equal bandwidth through the uplinks.

Calculate oversubscription

Suppose 40 servers each have a 25 Gb/s link. Their total potential demand is $40(25)=1000$ Gb/s. Four 100 Gb/s uplinks provide 400 Gb/s. The oversubscription ratio is

That is 2.5 : 1. If all 40 servers send equally to other racks and these uplinks are the bottleneck, their average is $400\mathrm{/}40=10$ Gb/s each. Traffic staying within the rack does not consume the uplinks in the same way. The ratio describes aggregate capacity, not an automatic fixed rate limit for every server.

Add paths with Clos networks

A Clos fabric provides multiple paths between racks using many smaller switches. For even k, a three-tier k-port fat tree has k pods, each with k/2 edge and k/2 aggregation switches, plus (k/2)² core switches. With equal-rate links it supports $k^3\mathrm{/}4$ hosts using $5k^2\mathrm{/}4$ switches.

For k = 4, the host count is $4^3\mathrm{/}4=16$ and the switch count is $5(4^2)\mathrm{/}4=20$. For k = 48, these become 27,648 hosts and 2,880 switches. The topology supplies full aggregate capacity for admissible traffic with suitable path assignment. It does not guarantee every host its full link rate when all hosts send to one destination or when routing concentrates flows.

Spread flows over equal-cost paths

Equal-cost multipath (ECMP) hashes packet fields such as IP addresses and transport ports to select a path. Packets of one flow normally follow the same path, preserving their order, while different flows can take different paths.

Hashing is not perfect load balancing. Two large flows may choose one path while another remains idle. Moving individual packets freely among paths can reorder them and trigger unnecessary TCP recovery. Capacity in the topology and capacity used by a particular traffic pattern are different measurements.

Modern designs also use software-defined control, VXLAN overlays for movable virtual machines and containers, and ECN-based congestion control for short queues and low latency. Operators balance bandwidth and resilience against the cost and energy use of switches, optics and links.

A new Ethernet connection must obtain addressing information before a browser can reach a remote site. Follow a laptop on 192.0.2.0/24, with gateway 192.0.2.1, DNS server 198.51.100.53 and web server 203.0.113.80. The exchange uses HTTPS over TCP.

Get the network settings

The laptop initially has a MAC address but no usable IPv4 address. It sends DHCP Discover from UDP port 68 to port 67, using IP source 0.0.0.0, IP destination 255.255.255.255 and Ethernet destination FF-FF-FF-FF-FF-FF.

The switch learns the laptop's source MAC and floods the broadcast. Here the router also runs the DHCP server. The four-message exchange is Discover, Offer, Request and Acknowledgment. The ACK supplies 192.0.2.37/24, gateway 192.0.2.1 and the DNS server address. A reply sent to the laptop's MAC can be forwarded using the switch's learned entry; no ARP lookup was needed for the initial broadcast.

Resolve the first next hop

The browser needs an IP address for www.example.com, so the laptop prepares a DNS query, normally to UDP port 53. The DNS server is outside the laptop's subnet. The laptop therefore needs the gateway's MAC address, not the DNS server's MAC.

With no cached gateway mapping, it broadcasts an ARP request for 192.0.2.1. The router replies with its local interface's MAC address. The laptop caches this mapping and puts the DNS datagram inside an Ethernet frame addressed to that interface.

Cross the routed path

The frame's destination is local, while its enclosed IP destination is 198.51.100.53. The campus router removes the Ethernet frame, looks up the IP route and builds the next link's frame. OSPF and BGP can supply the routes used within and between networks. Each subsequent router repeats next-hop forwarding.

DNS returns 203.0.113.80. The laptop can now create a transport connection to the web server. Its gateway MAC is already cached; reaching a different remote IP does not require resolving another local MAC.

Exchange encrypted web data

TCP connects to port 443 with SYN, SYN-ACK and ACK. A TLS handshake then establishes the encrypted channel. The browser sends an HTTP GET inside TLS over TCP. The server returns the page over that connection.

On the return path, the campus router places the final IP datagram in a frame addressed to the laptop. The switch forwards it to the learned port. The NIC checks the FCS, IP delivers to TCP, TCP orders the bytes, TLS processes the encrypted records and the browser reads the HTTP response.

Switches within one Ethernet LAN leave endpoint MAC addresses unchanged. Routers replace the frame for each next network; IP endpoint addresses remain unchanged in this trace. A NAT on another network could rewrite an IP address and port. Neither the outgoing frame nor the returning frame needs a switch port's MAC address as its destination.

A wireless network combines hosts, radio links and often a base station. Its channel changes as devices move, objects block paths and other transmitters become active. Those changes shape both transmission rates and recovery from errors.

Classify the connection

Infrastructure mode sends through a base station: a Wi-Fi access point or a cellular base station. Ad hoc mode has no such infrastructure. Either arrangement can use one wireless hop or several.

A Bluetooth central coordinates its piconet, but it is not an internet-access base station in this classification. Changing the serving AP or cellular base station is a handover.

Explain a weak or distorted signal

Path loss reduces received power with distance and obstacles. Interference adds unwanted energy from other transmitters; 2.4 GHz Wi-Fi shares spectrum with Bluetooth and microwave ovens. Multipath propagation brings reflected copies along different-length routes, which can blur symbols at the receiver.

The signal-to-noise ratio (SNR) compares received signal power with noise power:

A power ratio of 100 gives 20 dB. In a free-space model, power falls with distance squared. Doubling distance leaves one quarter of the power, a drop of about 6 dB if noise stays fixed. Indoors, walls and reflections also affect the received signal, so distance alone does not determine SNR.

Trade rate against errors

A modulation carrying more bits per symbol can send faster, but its signal states are harder to distinguish. For a fixed modulation, higher SNR generally lowers the bit error rate. Rate adaptation selects a more robust, slower mode when reception worsens and a faster mode when it improves. The highest advertised rate is not always the mode that delivers the most usable data under current conditions.

Locate the hidden collision

Suppose A and C can both reach B but cannot hear each other. A may hear silence while C is already sending to B. A then transmits and the frames interfere at B. This hidden terminal problem shows why a sender's carrier-sense result does not describe conditions at the receiver.

Classic Bluetooth serves short-range personal devices, often around 10 m, in the 2.4 GHz band. One central can coordinate up to seven active peripherals. It hops among 79 one-megahertz channels at 1,600 hops/s. Bluetooth Low Energy uses 40 two-megahertz channels. These are different physical channel arrangements, even though both support low-power personal devices.

Variable radio losses motivate local acknowledgments, retransmissions and FEC. A loss caused by weak reception can still reduce TCP throughput if transport recovery treats it as congestion.

Wi-Fi is the IEEE 802.11 wireless LAN family. An access point (AP) bridges wireless frames onto a distribution network, often Ethernet. It need not perform IP routing.

Distinguish a name from a radio

A basic service set (BSS) comprises an infrastructure AP and its associated stations. The SSID is the network name; the BSSID identifies a particular BSS, normally using the AP radio's MAC address. Several BSSs joined by a common distribution system and sharing an SSID form an extended service set (ESS). Two unrelated APs using the same name do not thereby form one ESS.

A station can find an AP by passive scanning, listening for periodic beacon frames, or active scanning, sending a probe request and listening for responses. Beacons advertise information such as the SSID and BSSID, typically about ten times a second. The station chooses a BSS and exchanges an association request and response. Strong signal is useful, but an AP with many competing clients may still deliver less throughput than a slightly weaker, quieter AP.

Put setup steps in order

Security and association are distinct operations. WPA2 first uses open 802.11 authentication and association. WPA2-Personal then establishes keys with the four-way handshake; WPA2-Enterprise performs an 802.1X credential exchange before that key handshake. WPA3-Personal instead uses SAE password authentication before association, followed by the four-way handshake. After successful link setup, DHCP can supply IP settings.

Compare bands and generations

The figures are peak physical rates using the relevant maximum capabilities, not each user's application throughput. Support for older generations requires a band and mode both devices share. Newer OFDMA features can schedule users, while contention remains central to Wi-Fi access.

Understand channel overlap

In the ordinary 2.4 GHz channel sequence, centers are 5 MHz apart. Channels 1, 6 and 11 are centered at 2,412, 2,437 and 2,462 MHz. A roughly 20 MHz signal overlaps close neighbors: channels 1 and 4 are only 15 MHz apart, while 1 and 6 are 25 MHz apart.

The 5 and 6 GHz bands offer more separate channels. Channel bonding joins spectrum into 40, 80 or 160 MHz channels; Wi-Fi 7 also supports 320 MHz channels. Available channels and permitted widths depend on local radio regulations, especially at 6 GHz. Wider channels can increase rate, but the advertised width does not create dedicated spectrum for each station.

Wi-Fi uses CSMA/CA, carrier sensing with collision avoidance. A transmitter's own radio signal overwhelms weak incoming signals, and hidden stations may be inaudible. It sends a complete frame and learns success from an acknowledgment.

Count down only through idle slots

In the basic contention procedure, a station waits for an idle DIFS interval. If it has deferred, it chooses a random backoff and counts down during idle slots. A busy channel freezes the remaining count. After the exchange ends and DIFS passes, countdown resumes.

If X chooses 3 and Y chooses 7, X reaches zero first. Y freezes with 4 slots remaining while X transmits. A correct unicast data frame receives an ACK after the shorter SIFS interval. SIFS gives that response priority over stations waiting DIFS. Y resumes after the exchange; it does not restart automatically at 7 or count through X's transmission.

A missing ACK causes a retry with a larger contention window. It can mean lost data or a lost ACK. Sequence numbers let the receiver suppress a duplicate in the latter case. Ordinary broadcast and multicast data frames receive no individual ACKs.

Reserve airtime for hidden senders

Optional RTS/CTS exchanges reserve a duration before a long data frame. A sender sends request-to-send; the receiver replies clear-to-send. A hidden station that hears the CTS defers for the announced time, even if it never heard the RTS. Short control frames add overhead but can avoid wasting a long data transmission in a collision. They reduce hidden-terminal damage without making the channel collision-free.

Follow three address roles

A basic data frame carries frame control, duration, addresses 1–3, sequence control, an optional address 4, a payload and a 32-bit CRC, in that order. The traditional maximum payload is 2,312 bytes; newer aggregation formats add further structure. A fourth address appears in four-address formats such as wireless bridging. In ordinary infrastructure data frames, address 1 is the immediate radio receiver and address 2 the radio transmitter. Address 3 preserves the endpoint on the distribution-system side.

For host H1 sending through AP to router interface R1:

1. Toward AP: address 1 = AP, address 2 = H1, address 3 = R1.
2. AP emits Ethernet with source H1 and destination R1.
3. In return, Ethernet has source R1 and destination H1.
4. AP sends radio data with address 1 = H1, address 2 = AP, address 3 = R1.

Address 3 can belong to another LAN host instead of a router. The AP's radio role must not replace the original Ethernet endpoint in the bridged frame.

Move or sleep

Moving from AP1 to AP2 on the same subnet preserves H1's IP address. Reassociation changes its radio attachment. A frame arriving through AP2 with source H1 teaches switches H1's new location. Movement to a different subnet needs new addressing or mobility support to preserve existing connections.

For power saving, a station tells the AP it will sleep. The AP buffers its frames and advertises waiting traffic in beacon information. The station wakes to check and collect data, then can sleep again. Sleeping saves radio energy but can delay delivery until the next wake interval.

A cellular network divides coverage into cells served by base stations. The phone or modem is user equipment (UE). Its SIM holds a subscriber identity, the IMSI, and a long-term secret used in authentication.

Separate network roles

An IMSI has up to 15 digits: a three-digit mobile country code, a two- or three-digit network code, and a subscriber number. The test identity 001 01 0123456789 illustrates those fields. It identifies a subscription, not the UE's IP address.

The LTE base station, an eNodeB, schedules radio resources, handles its cell and participates in handover. Its backhaul leads to the Evolved Packet Core (EPC).

Control messages establish and manage connectivity. User-plane packets carry the subscriber's actual data. Treating the MME as the router for every web packet confuses those planes.

Attach step by step

The UE finds synchronization signals and selects a cell. It sends an attach request through the eNodeB to the MME. The MME obtains authentication information from the subscriber's home HSS. The UE and network authenticate and establish protection, then the MME coordinates the bearer data paths through the gateways. The P-GW provides the UE's IP address.

LTE is an all-IP packet system; supported voice service travels as IP as well as ordinary data. Between the eNodeB, S-GW and P-GW, GTP-U carries the UE's packets through tunnels over UDP and outer IP. The inner datagram retains the UE endpoint address while outer addresses identify tunnel endpoints. Changing a radio attachment can therefore change a tunnel endpoint without changing the UE's IP address.

Divide the radio work

The link stack has three main sublayers. PDCP performs functions including header compression and encryption. RLC segments and reassembles data and can provide retransmission in acknowledged mode. MAC manages scheduled opportunities and hybrid ARQ, combining rapid retries with coding.

LTE's downlink uses OFDMA; the uplink uses the related SC-FDMA scheme. Subcarriers are 15 kHz apart. A resource block spans 12 subcarriers, or 180 kHz, for a 0.5 ms slot. Scheduling commonly assigns resources in 1 ms subframes. More allocated blocks and a higher usable modulation rate can increase a UE's data rate; the standard does not require one particular fairness algorithm.

An idle UE saves energy using sleep and wake procedures. Only the access hop must be radio: backhaul can be fiber, copper or microwave, and gateways still separate the radio network from the internet.

5G extends cellular access with flexible radio timing and a redesigned core. High peak rates, low latency and large device populations serve different use cases; one device should not expect all design targets at once.

Interpret the design goals

IMT-2020 defines enhanced mobile broadband, eMBB; ultra-reliable low-latency communication, URLLC; and massive machine-type communication, mMTC. Associated requirements include a 20 Gb/s peak downlink rate, 1 ms user-plane radio latency for URLLC, and one million devices per square kilometer for dense machine communication. These are evaluation targets under specified conditions, not ordinary internet download speed or round-trip latency guarantees.

Change the radio timing

New Radio (NR) uses OFDM with scalable spacing. With numerology μ, subcarrier spacing is $15(2^{\mu })$ kHz and a normal 14-symbol slot lasts $1\mathrm{/}{2}^{\mu }$ ms. At 30 kHz, μ = 1 and the slot lasts 0.5 ms. At 120 kHz, μ = 3 and it lasts 0.125 ms. Shorter slots enable faster scheduling opportunities, though end-to-end latency also includes transport, processing and queues.

NR FR1 spans 410 MHz–7.125 GHz, including common mid-band around 3.5 GHz. FR2 spans 24.25–71 GHz across its subranges. The high bands often called millimeter-wave offer wide channels but suffer strong blockage and short useful range. Small cells and focused beams from large antenna arrays help use them. Many 5G deployments use FR1 and need no millimeter-wave link.

Map core responsibilities

A gNB is a 5G base station. Non-standalone (NSA) deployments combine NR access with an LTE anchor and 4G core. Standalone (SA) deployments use the 5G core's service-based organization.

The functions divide responsibilities; the mapping is not one old box replaced by one identically functioning new box. Separating user-plane forwarding from control lets an operator place a UPF near a local application. Network slicing supports virtual networks with different service properties on shared infrastructure.

Compare cellular and Wi-Fi access

Cellular networks generally use licensed spectrum, operator-managed subscribers and centrally scheduled radio resources over wide areas. Wi-Fi generally uses unlicensed local spectrum and contention-based access managed by an organization or user. Wi-Fi 6 and later also support AP-scheduled OFDMA, so scheduling is not unique to cellular. Both families adapt radio rates and recover some losses locally. 5G-Advanced continues the 5G family; 6G remains a development effort aimed toward around 2030.

Changing radio attachment is easy to hide if a device keeps its IP address. Moving between networks is harder: an ordinary TCP connection is identified by endpoint addresses and ports. Mobility mechanisms preserve reachability while the physical delivery path changes.

Separate stable identity from current location

The home network holds the subscription record; a visited network serves a roaming subscriber elsewhere. In Mobile IP, a stable home address identifies the device and a care-of address locates its current attachment. Advertising a separate route for every moving device throughout the internet would create an enormous, constantly changing routing burden.

With indirect routing, the correspondent sends to the home address. A home agent intercepts and tunnels traffic toward the care-of address. The correspondent need not track movement, but the path can detour through the home network. A Brazilian subscriber in Tokyo receiving data from a Tokyo server may have that traffic sent through Brazil first: triangle routing.

With direct routing, the correspondent learns the current location and tunnels toward the visited network. It avoids the home detour but requires correspondent participation and updates or forwarding when the device moves again. In LTE roaming, home routing through a home P-GW and local breakout provide related choices.

Follow an LTE handover

Within an LTE network, the S-GW can remain the data-path anchor while the UE changes eNodeBs:

1. Signal measurements reported by the UE lead the source eNodeB to request handover from a target.
2. The target reserves resources and returns the access information.
3. The source commands the UE to move and forwards in-flight data toward the target.
4. The UE synchronizes with the target, accesses it and sends completion.
5. The target requests a path switch through the MME; the S-GW changes its tunnel endpoint.
6. The old radio resources are released, and traffic follows the new tunnel.

Preparation happens before the UE moves; the core path switch follows arrival at the target. A handover may briefly delay or lose packets without changing the inner IP address.

Distinguish Mobile IP versions

Mobile IPv4 supports a foreign-agent care-of address or one obtained directly by the mobile device. Discovery and registration inform the home agent. Mobile IPv6 needs no foreign agent: ordinary IPv6 configuration obtains the care-of address, and a Binding Update registers it with the home agent. These IETF mechanisms are less widely deployed than cellular-specific mobility management.

Connect losses to transport behavior

Conventional loss-based TCP generally treats loss as congestion, so a radio error or handover loss can reduce its sending rate even when queues were not overloaded. Local ARQ and FEC can hide some wireless losses. Other approaches distinguish loss causes or split a transport connection across wireless and wired portions. Splitting changes the original end-to-end transport arrangement. Applications such as streaming video can also lower their bitrate as available bandwidth falls.