Physical and data link basics

Why it matters

Two numbers set the limits for every network design above them: how many bits per second a link can carry, and how long a bit takes to arrive. Neither can be fixed in software. The speed of light is why a cross-ocean request takes at least 60 milliseconds no matter how fast your servers are, and why multi-region systems are designed the way they are. The link layer, meanwhile, decides what happens on the last hop: Wi-Fi's shared air is why a laptop's latency jitters when a phone fetches its backups.

The idea

The physical layer turns bits into signals (voltage levels on copper, light pulses in fibre, radio waves in the air) and back. The data link layer groups bits into frames, addresses them to a neighbour on the same link, detects corruption, and decides who may transmit when the medium is shared.

Bandwidth, latency and the speed of light

Bandwidth is how many bits per second a link carries; latency is how long one bit takes to get there. They are independent: a truck full of hard drives has huge bandwidth and terrible latency. Latency on a path has four parts.

Component	What it is	Typical size
Propagation	Distance divided by signal speed (about 200,000 km/s in fibre, two thirds of light in vacuum)	~5 µs per km: ~25 ms across the US, ~35 ms London–New York one way
Transmission (serialisation)	Packet size divided by link bandwidth	1,500 B on 1 Gbit/s: 12 µs; on 10 Mbit/s: 1.2 ms
Queuing	Waiting in router and switch buffers behind other packets	Zero to hundreds of milliseconds; the variable part
Processing	Routers reading headers and choosing a path	Microseconds per hop

Propagation is the floor. A London to Sydney round trip cannot be shorter than about 160 ms over fibre, so a request that needs three sequential round trips to a far region costs half a second before any server does any work. This is why CDNs, edge caching and regional deployments exist.

Framing and error detection

A frame wraps a payload with a header (source and destination MAC address, type of payload) and a trailer holding a CRC checksum. The receiver recomputes the CRC; if it differs, the frame is silently dropped. The link layer detects errors but on wired Ethernet does not retransmit; recovering is TCP's job higher up. Wi-Fi, where errors are frequent, does acknowledge and retransmit frames at the link layer, which hides loss from TCP at the cost of variable latency.

Sharing the medium

Medium	How it shares	Consequence
Switched Ethernet (full duplex)	Each device has a dedicated cable to a switch port; both directions at once	No collisions; predictable latency
Old shared Ethernet (hubs)	Everyone on one wire; detect collisions and back off randomly	Historical; why 'collision domain' is in old textbooks
Wi-Fi	One radio channel shared by every device and neighbour network; listen before talking, back off on contention	Throughput divides among active devices; latency jitters; distance and walls reduce the rate
Cellular (4G, 5G)	The base station schedules who transmits when	Higher and more variable latency than wired; radio wake-up adds tens of milliseconds

Physical realities that reach application engineers

Mobile clients are often on links with 50 to 200 ms latency, variable bandwidth and loss, so designs that need many round trips feel slow on phones even when they are fast in the office. Data centre links are 10 to 400 Gbit/s with sub-millisecond latency, so inside a region the bottleneck is usually the servers, not the network. Cross-region links are where propagation dominates, which is why replication between regions is usually asynchronous.

Trade-offs

Higher bandwidth is buyable; lower propagation latency is not, except by moving closer. Wired links give predictable performance at the cost of mobility; wireless gives mobility at the cost of shared, variable capacity. Link-layer retransmission (Wi-Fi) smooths loss for TCP but adds latency variance; leaving recovery to TCP keeps links simple but makes loss visible end to end.

In an interview

Use propagation to justify architecture: "Users in Sydney are 160 ms round trip from us-east, and a cold HTTPS request needs three round trips, so we terminate TLS and cache at an edge location near them and keep only the necessary request crossing the ocean." For mobile: "Every round trip costs 100 ms on cellular, so the app makes one batched call instead of five sequential ones."