The Network Edge & Core
The Internet is easiest to reason about when you split it into two regions: the edge, where the things you actually use live, and the core, the mesh of switches and links that carries traffic between them.
The network edge
The edge is made of end hosts (also called end systems): laptops, phones, servers, sensors - anything that runs applications and sits at the "end" of the network. End hosts connect to the rest of the Internet through an access network, the first hop out of your building or device:
- Home access - cable, DSL, or fiber to the home (FTTH), usually terminating at a modem and router.
- Cellular access - 4G/5G radio links to a nearby base station, then into the carrier's network.
- Enterprise / institutional access - Ethernet and Wi-Fi inside an office or campus, aggregated into a link to a provider.
The defining trait of the edge is that hosts originate and consume data. They do not, as a rule, forward other people's traffic.
The network core
The core is a mesh of packet switches - mostly routers - connected by links (fiber, microwave, undersea cable). A router's whole job is to receive a packet on one link and forward it out another, moving it one hop closer to its destination.
How a packet traverses the core
Your data is chopped into packets, each stamped with a destination address. Each router along the way independently looks at that address, consults its forwarding table, and sends the packet out the appropriate link. No router knows the full path; it only knows the next hop. The path emerges hop by hop.
Packet switching vs circuit switching
There are two fundamentally different ways to move data through a shared network. Circuit switching (the old telephone model) reserves a dedicated path with guaranteed capacity for the whole conversation. Packet switching (the Internet model) sends independent packets that share links on demand.
| Property | Circuit switching | Packet switching |
|---|---|---|
| Resource allocation | Reserved up front for the whole call | On demand, per packet |
| Path | Fixed dedicated circuit | Each packet routed independently |
| Efficiency under bursty traffic | Low - idle reservation is wasted | High - links shared statistically |
| Guaranteed rate | Yes | No (best effort) |
| Behavior under overload | New calls blocked | Queuing delay and possible loss |
| Example | Traditional phone network | The Internet |
Store-and-forward
A packet switch must receive the entire packet before it can start transmitting it onto the next link - it stores the whole packet, checks it, then forwards it. That is why transmission delay is paid once per link on the path: a packet of L bits crossing N links each at rate R takes at least N x L / R seconds just in transmission, before any propagation or queuing.
Worked example: store-and-forward delay across a path
Take a 2-link path carrying a 5000-bit packet: the first link runs at 1 Mbps, the second at 100 Mbps. Because each switch must store and forward the whole packet, the transmission delay is paid once per link, in sequence, not overlapped:
- Link 1 transmission delay: 5000 bits / 1,000,000 bps = 5 ms.
- Link 2 transmission delay: 5000 bits / 100,000,000 bps = 0.05 ms.
The packet cannot start going out on link 2 until the switch has received every bit of it over link 1, so the two delays add: 5.05 ms of pure transmission delay before propagation or queuing are even considered. Notice that the slow first link (1 Mbps) dominates the total almost completely - a single slow hop near the edge can erase the benefit of a fast core link further along the path, which is the same bottleneck idea from throughput, now expressed as delay instead of rate.
Statistical multiplexing
Because packets are forwarded on demand rather than on a reservation, many flows can share the same link, each using it only when it has something to send. This statistical multiplexing is why packet switching is so efficient for bursty traffic: the link is filled by whoever needs it at that instant, not left idle by an absent reservation holder. The cost is that when too many packets arrive at once, they wait in a queue - and if the queue fills, packets are dropped.
The ISP tier hierarchy
No single network reaches everywhere, so the core is really a network of networks stitched together by Internet Service Providers arranged roughly in tiers:
- Access ISPs - the providers your edge connects to directly (your home or mobile carrier).
- Regional ISPs - larger networks that aggregate many access ISPs across a region and carry their traffic upward.
- Tier-1 ISPs - global backbone networks that reach each other without paying anyone for transit; they form the top of the hierarchy.
Internet Exchange Points
Rather than always routing traffic up through higher tiers, networks meet directly at Internet Exchange Points (IXPs) - shared facilities where many ISPs and content providers interconnect and peer, exchanging traffic directly. Peering at an IXP shortens paths, cuts transit cost, and keeps local traffic local instead of hauling it across the backbone.
Large content providers (streaming, search, social) increasingly connect directly to access ISPs at IXPs, so the classic strict tier hierarchy is now flatter than the textbook picture.
Peering is a business relationship, not a technical guarantee, and when it breaks down customers can lose reachability outright. In 2008, a commercial dispute between two large backbone networks, Cogent and Telia, led both sides to depeer - simply unplugging the direct link between them - because neither would pay the other for transit. For days, customers of one network could not reach customers of the other at all; traffic had no alternate path because both networks refused to buy transit through a third party to route around the dispute. The lesson for anyone depending on Internet reachability for a business: a single upstream provider, however large, is still a single point of failure if a peering relationship it depends on collapses.