Networking & Protocols
The physical link: multiple-choice review
Crux Multiple-choice synthesis across the physical-link unit — latency floors, Shannon ceilings, bufferbloat, RoCE fabrics, and the frontier.
Your altitude — climbing toward senior
ZeroJuniorMiddleSenior
You are at senior altitude — in orbitSix questions that cut across the whole unit. Each one mirrors a call you make in a real incident — distinguishing physics from a bug, picking the right lever, reading the failure mode — not a definition to recite.
Goal
Confirm you can connect propagation floors, the Shannon ceiling, bandwidth-delay product, bufferbloat versus long-RTT congestion control, and the datacentre fabric — the synthesis the individual lessons built toward.
Quiz
Completed
A team upgrades its NYC→Sydney link from 10 Gbps to 100 Gbps and is surprised that page-load latency does not improve. What is the correct read?
Heads-up A bandwidth upgrade takes effect on the existing path immediately; no restart changes the speed of light over 16,000 km.
Heads-up Faster line rates lower serialisation delay, not raise it. Serialisation is microseconds here regardless; the 160 ms floor is propagation, not serialisation.
Heads-up A small window does cap throughput on a high-BDP path, but that is a throughput problem — it still does not change the propagation floor that sets first-byte latency.
Quiz
Completed
An RF team wants more Wi-Fi capacity. They can either double the channel width or push from 1024-QAM to 4096-QAM. Which moves capacity more, and why?
Heads-up QAM density rides the logarithmic S/N term: 1024→4096-QAM adds only 2 bits/symbol and needs ~45 dB SNR, which most real links never sustain. Bandwidth is the linear lever.
Heads-up They are not equivalent. Doubling B doubles capacity; the QAM jump is a small logarithmic gain that often regresses under real-world noise.
Heads-up Both B and modulation order are exactly the levers the standard exposes; SNR decides which QAM is even achievable.
Quiz
Completed
A household runs a video call that stutters whenever a cloud backup uploads, yet the speed test reads 300/40 Mbps green. Idle ping is 18 ms; under upload it climbs to 420 ms. Diagnosis and first fix?
Heads-up Throughput was never the problem; the speed test is green. A faster uplink raises the saturation ceiling but leaves the oversized buffer, so ping still balloons under the backup.
Heads-up The bloated buffer sits at the edge modem, upstream of the LAN. Wired or wireless, the modem queue still fills under saturation; wiring does not touch it.
Heads-up A lower-bitrate codec sends fewer bytes but still queues behind the full buffer. The latency comes from queueing delay, not from the call's bandwidth.
Quiz
Completed
The same loss-based CUBIC sender runs well over a Starlink LEO link (~50 ms RTT) but starves over a GEO link (~600 ms RTT). Why, and what is the right response on GEO?
Heads-up The problem is the length of the feedback loop, not a high loss rate. CUBIC starves because its loss signal is stale by the time it arrives, not because losses are frequent.
Heads-up Capacity is not the issue — CUBIC underutilises the GEO pipe it has. The fix is a loss-independent algorithm (BBR), not more bandwidth.
Heads-up AQM addresses bufferbloat (a queue problem), not a slow control loop. On GEO the issue is the algorithm reacting too late; BBR is the lever, not queue management.
Quiz
Completed
A 64-GPU training job runs 3× slow. Each leaf has 2×400G uplinks but 8×200G server NICs, and PFC PAUSE frames appear on every uplink. What is happening?
Heads-up No-drop is exactly the point of PFC, but the PAUSE frames are the symptom of congestion, not health. GPU utilisation is low because the oversubscribed fabric stalls the collective.
Heads-up RoCE deliberately bypasses the kernel TCP stack for zero-copy RDMA; it requires lossless Ethernet (PFC/ECN). The fix is correct bisection bandwidth, not switching transport.
Heads-up Disabling PFC makes the fabric lossy; RDMA has no retransmission, so dropped packets stall or abort the collective entirely. PFC is required — the fix is removing the oversubscription causing the backpressure.
Quiz
Completed
A high-frequency-trading firm wants the lowest possible NYC↔Chicago round-trip latency. Which physical-layer investment actually helps?
Heads-up Coherent 800G raises throughput, but propagation delay is set by path length and signal speed. A bandwidth upgrade does not move the latency floor — the wrong lever for a latency-bound goal.
Heads-up A ~550 km up-and-down LEO hop adds round-trip distance versus a direct ~1,200 km terrestrial route. LEO is for coverage and redundancy, not for beating fibre between two well-connected cities.
Heads-up Post-quantum crypto changes the handshake's security posture, not its propagation delay. Right call for long-term confidentiality, irrelevant to a latency objective.
Recap
The through-line of the unit is one separation: the propagation floor (distance ÷ signal speed) is physics you design around, and everything above it is engineering you can fix. Shannon caps a link’s bits/s (bandwidth linear, SNR logarithmic); BDP says how many bytes must be in flight to fill it; bufferbloat and long-RTT starvation are two different congestion failures with two different fixes (AQM versus BBR); the datacentre fabric trades the same bandwidth and loss budgets at hyperscale; and only a shorter path — or hollow-core fibre — ever lowers the floor.