Chapter 6 Review · The Link Layer and LANs

Link layer ★ Chapter review Kurose & Ross pp. 518–519 · ~13 min read

  • link
  • multiple access problem
  • switch table

Where you are

  • Application layer
  • Transport layer
  • Network layer
  • Link layer you are here
  • Physical layer

This chapter was the bottom of the journey: not how a packet crosses the world, but how it crosses one wire with other machines shouting on it.

What this page is

  • The chapter in one picture — every idea, linked to the section that taught it.

  • The argument in five steps, which is how the chapter hangs together.

  • The three multiple access families, scored side by side.

  • Every number worth remembering, sortable.

  • The book’s slips in this chapter — five confirmed, one open.

  • A closing quiz on the chapter as a whole.

Where you are

The link layer ’s basic service is small: move a datagram from one node to an adjacent node. Every link-layer protocol does it by encapsulating that datagram in a frame.

Beyond that shared framing job, though, the protocols differ enormously — because the links do. A point-to-point link has one sender and one receiver. A multiple access link is shared, and needs a protocol just to decide who may speak.

The journey down the protocol stack is now over. The physical layer lies below, but its details belong to a course in communication theory rather than computer networking.

The chapter in one picture

Chapter 6 in one picture

Hover an idea to see what it connects to. Click one to go back to the section that taught it.

Every idea, linked to the section that taught it. Click any box.

Read this map as a list
  • Services and the adapter (§6.1) — connects to Error detection, Channel partitioning, MAC addresses and ARP
  • Error detection (§6.2 · parity, checksum, CRC) — connects to Services and the adapter, Ethernet
  • Channel partitioning (§6.3.1 · TDM, FDM, CDMA) — connects to Services and the adapter, Random access: ALOHA
  • Random access: ALOHA (§6.3.2 · 1/e and 1/(2e)) — connects to Channel partitioning, CSMA and CSMA/CD
  • CSMA and CSMA/CD (§6.3.2-csma · sense and abort) — connects to Random access: ALOHA, Taking turns, and DOCSIS
  • Taking turns, and DOCSIS (§6.3.3–6.3.4) — connects to CSMA and CSMA/CD
  • MAC addresses and ARP (§6.4.1) — connects to Services and the adapter, Ethernet, A day in the life
  • Ethernet (§6.4.2 · the frame that never changed) — connects to MAC addresses and ARP, Switches, self-learning, Error detection, A day in the life
  • Switches, self-learning (§6.4.3) — connects to Ethernet, VLANs and 802.1Q, MPLS, A day in the life
  • VLANs and 802.1Q (§6.4.4) — connects to Switches, self-learning
  • MPLS (§6.5 · traffic engineering) — connects to Switches, self-learning, Data centres
  • Data centres (§6.6 · Clos networks) — connects to MPLS
  • A day in the life (§6.7 · all of it, running at once) — connects to MAC addresses and ARP, Ethernet, Switches, self-learning

The argument, in five steps

1 — A frame is not a datagram, and an adapter is not software

Section 6.1 put the link layer in hardware: the network adapter. That is not a detail. It is why the CRC (Cyclic Redundancy Check) can be more expensive than a transport-layer checksum, and why an adapter can transmit and listen at the same time.

Section 6.2 then showed what those bits buy: parity detects, two dimensional parity corrects one error, and CRC-32 catches every burst shorter than 33 bits.

2 — Sharing one channel is a genuinely hard problem

Four properties are wanted, from section 6.3.1: full rate for one active node, a fair share for many, decentralisation, simplicity.

No family gets all four. Partitioning fails the first, random access fails the second, taking turns fails the third. The chapter’s honesty is in refusing to name a winner.

The three multiple access families, scored
Channel partitioning§6.3.1Random access§6.3.2Taking turns§6.3.3
One active node gets the full rate R
M active nodes each get about R/M
No single point of failure
Best efficiency reached
Who uses it

Cells marked ⓘ have a reason behind them — click to read it.

Section 6.3’s whole argument. No column is all good — this is a set of trade-offs, not a ranking. Click any cell.

3 — Listening is worth a factor of three

Slotted ALOHA never listens, and cannot pass 37 per cent. Add carrier sensing and collision detection and a real Ethernet reaches about 98 per cent.

But sensing does not prevent collisions, and section 6.3.2-csma is built around why. A node senses the channel only where it is standing, so “idle” is always older news than “idle everywhere” — by exactly the propagation delay.

4 — Two addressing schemes, doing different jobs

A MAC (Media Access Control) address is flat and permanent; an IP (Internet Protocol) address is hierarchical and changes when you move. One says who you are, the other where.

So a datagram crossing a router needs two frames and four MAC addresses. §6.4.1’s capture proves it: the same datagram on either side of a router, with identical IP addresses and no Ethernet address in common.

Then switches made the collisions go away entirely. Section 6.4.2 ends by admitting that in a switched, full-duplex Ethernet there is no need for a MAC protocol at all.

5 — A whole network can be a link

MPLS (Multiprotocol Label Switching) 6.5) and cable access (§6.3.3-6.3.4) both do it, and so does a data centre fabric (§6.6).

The book says this should not seem odd: a telephone link from a home modem to a remote router is already a path through a large and complicated network.

Every number worth remembering

Every number worth remembering
NumberWhat it isSection

Cells marked ⓘ have an explanation — click to read it.

Sortable. Each was checked against the book during the build; several are asserted by a tool on every run.

The book’s slips in this chapter

The book’s slips in this chapter
WhereWhat it saysWhy it is wrong

Cells marked ⓘ have an explanation — click to read it.

Five confirmed and one open question, every one checked against the printed page rather than the text dump. They are here because a reader who hits one should know it is the book and not them.

In plain words

Chapter 6 has more errata than any chapter so far, and the reason is visible in the table: three of the six are edition-update damage. Numbers were changed for the 8th edition and the sentences around them were not.

The Figure 6.27 one is the most instructive. A typo in a figure annotation would normally be trivial. But that annotation is the only place in the whole section where the book mentions that the frame check has to be recomputed.

Two of these were found only because every figure on this site is transcribed from the page image rather than from the text dump. The dump drops figure annotations entirely.

Check yourself

Check yourself — the whole chapter

0 of 8 answered
  1. 1.What is the link layer's basic service?

  2. 2.Section 6.3 lists four properties a good multiple access protocol should have. How do the three families score?

  3. 3.predictA datagram crosses one router on its way from one host to another. How many frames and how many MAC addresses are involved?

  4. 4.Carrier sensing means a node listens before transmitting. Why do collisions still happen?

  5. 5.A switch receives a frame. Which address does it store in its table, and which does it look up?

  6. 6.Section 6.4.2 ends by asking whether Ethernet still needs a MAC protocol. What is the answer?

  7. 7.What do MPLS, a cable access network and a data centre fabric have in common?

  8. 8.predictIn §6.7's capture, Bob's laptop performs exactly one ARP exchange. Why only one?

The chapter in three lines

  • The link layer moves a datagram to an adjacent node, by putting it in a frame. That is all it promises — and it lives in hardware, which is why it can afford a CRC and can listen while transmitting.
  • Sharing a channel has three families of answer and no winner among them. Listening before and during transmission is worth roughly a factor of three over ALOHA.
  • MAC addresses are flat, IP addresses are hierarchical, and ARP (Address Resolution Protocol) translates — but only within one subnet. Switches learn from source addresses, decide from destination addresses, and configure themselves.