🌐 Computer Networks

What you’ll learn

1. Nobody Is In ChargeEstablish the end-to-end principle as the internet's founding decision, and see that the network's ignorance is deliberate.The telephone network was smart in the middle and dumb at the edges; the internet inverted that. Saltzer, Reed and Clark's 1984 end-to-end argument holds that functions needing endpoint knowledge should live at the endpoints, so the network promises only best effort — it may drop, duplicate, delay or reorder your data, and guarantees nothing. That refusal to promise is what let anything be built on top without permission, and it is also the root of the fragilities in later chapters.
2. Why Packets, and Not CircuitsUnderstand statistical multiplexing — why chopping data into independently-addressed packets beats reserving a path, and what it costs.A reserved circuit is idle most of the time because data traffic is bursty, and reserving capacity for everyone's peak means buying capacity for a peak that never happens simultaneously. Packet switching instead chops data into independently-addressed chunks that share every link, so idle moments are instantly used by someone else. The price is that packets can be dropped, reordered or duplicated, and that congestion is now possible — which is what chapters 8 and 9 exist to clean up.
3. Layers: Deliberate IgnoranceUnderstand layering as enforced ignorance, and see why that ignorance is what let the internet survive every technology change under and over it.Each layer uses the one below without knowing how it works and serves the one above without knowing what it is carrying. IP is the narrow waist: everything above runs over IP, and IP runs over everything below. That is why Wi-Fi could be invented without changing the web, and the web without changing Ethernet. The seven-layer OSI model is the teaching vocabulary; the four-or-five-layer TCP/IP stack is what was actually built.
4. Two Kinds of AddressUnderstand why the internet needs both a permanent hardware name and a hierarchical location — and why the difference is not redundancy.A MAC address is a 48-bit factory-assigned name that says who a device is but nothing about where it is; an IP address is a location within a hierarchy, which is the only reason routing can scale. Because IPv4's 32 bits allow about 4.3 billion addresses, the world ran out; IPv6's 128 bits are enough that nobody will. NAT bought decades of time by letting a household share one public address, at the cost of quietly breaking the end-to-end principle.
5. Switching and Routing Are Two Different JobsSeparate the local delivery problem from the global one, and see that the internet's core is a chain of individually ignorant decisions.A switch delivers within one local network using MAC addresses and a table it learns by watching traffic; a router moves packets between networks using IP prefixes. The crucial property is that no router knows the whole path — each one knows only the next hop, so a packet's journey is assembled from a chain of independent local decisions, with the TTL field as the safety net that stops loops from lasting forever.
6. BGP: The Internet Runs on TrustSee how independent networks agree on routes, and understand why the protocol that glues the internet together is built on assertion rather than proof.The internet is tens of thousands of independent autonomous systems that must agree on reachability without any authority. BGP is how they do it: each network announces which prefixes it can reach and neighbours propagate that. But an announcement is an unverified claim, and combined with longest-prefix-match it means a false, more-specific announcement wins — as when Pakistan Telecom's 2008 attempt to block YouTube domestically pulled YouTube's traffic worldwide for up to two hours. RPKI is a real, partial fix.
7. DNS: The Layer of NamesUnderstand how names become addresses through delegated hierarchy and caching, and why this layer is a chokepoint the rest of the design tried to avoid.Routing needs addresses; humans need names. DNS bridges them with a delegated tree — the root knows only who runs .com, .com knows only who runs each domain — so no server holds the whole database and authority is distributed. Thirteen root addresses exist because a priming response had to fit in a 512-byte UDP packet; anycast turns those into roughly 1,900 physical instances. Caching is what makes it survive its load, and TTL is the dial between speed and freshness.
8. TCP: Reliability Built at the EdgesSee the end-to-end principle made concrete — how two endpoints construct a reliable ordered stream on top of a network that guarantees none of it.TCP numbers every byte, acknowledges what arrives, retransmits what is not acknowledged, and reorders at the receiver — turning a lossy unordered packet service into a reliable ordered stream, entirely at the endpoints. UDP deliberately declines all of that, which is the right choice for real-time media where a late packet is worthless. QUIC re-implements TCP's guarantees over UDP to escape TCP's own drawbacks, including head-of-line blocking and the difficulty of changing anything ossified into middleboxes.
9. Congestion Control: The Unwritten TreatyUnderstand the internet's most remarkable property — that it is held together by voluntary restraint at the endpoints, and what happened the one time it failed.In October 1986 throughput between two sites 400 yards and three hops apart collapsed from 32 kbit/s to 40 bit/s: senders responded to loss by retransmitting, which caused more loss. Jacobson and Karels's answer treats loss as the network's only congestion signal and has senders back off multiplicatively and probe additively — a rule with no enforcement anywhere. Every device on the internet voluntarily slows down when it detects loss, and the internet works because almost all of them keep that treaty.
10. One Page, Every LayerAssemble the whole course into the two seconds after you press Enter, and land the through-line.Loading a page runs the entire stack: DNS resolves the name through delegated caches, a transport handshake establishes a connection, TLS negotiates encryption, HTTP requests the document, and hundreds of packets are routed hop by hop by machines that know only their next step while congestion control continuously negotiates rate with a network that never speaks. None of it is coordinated by anyone, and the whole thing completes before you notice.

Questions this course answers

Grounded in trusted sources

• J. H. Saltzer, D. P. Reed, D. D. Clark, 'End-to-End Arguments in System Design', ACM TOCS, 1984 — https://web.mit.edu/Saltzer/www/publications/endtoend/endtoend.pdf
• Van Jacobson and Michael J. Karels, 'Congestion Avoidance and Control', ACM SIGCOMM, 1988 — https://ee.lbl.gov/papers/congavoid.pdf
• ESnet, 'Unjamming the Information Superhighway and saving the Internet' (the October 1986 collapse: 32 kbit/s to 40 bit/s between LBL and UC Berkeley) — https://www.es.net/about/esnet-history/unjamming-the-information-superhighway-and-saving-the-internet/
• RIPE NCC, 'YouTube Hijacking: A RIPE NCC RIS case study' (24 February 2008) — https://www.ripe.net/about-us/news/youtube-hijacking-a-ripe-ncc-ris-case-study/
• root-servers.org — root server instance counts (approximately 1,950 instances as of December 2025) — https://root-servers.org/
• J. Postel (ed.), RFC 791, 'Internet Protocol', IETF, 1981 — https://www.rfc-editor.org/rfc/rfc791
• J. Postel (ed.), RFC 793, 'Transmission Control Protocol', IETF, 1981 — https://www.rfc-editor.org/rfc/rfc793
• P. Mockapetris, RFC 1035, 'Domain Names — Implementation and Specification', IETF, 1987 — https://www.rfc-editor.org/rfc/rfc1035

Every Wunder lesson is built from real, reputable sources — never invented.

Related Science courses