Nothing in reliability or ordering tells the sender how fast to send. Two separate things can be overwhelmed, and holding them apart is the whole lesson.
Flow control protects the receiver. Data lands in a buffer awaiting read(); a fast server streaming to a slow phone overflows it. But the receiver knows its free space and is already sending ACKs, so the fix is nearly free: advertise free buffer space in every ACK — the receive window (rwnd). A zero window stops the sender precisely, without loss.
Congestion control protects the network. A receiver advertising 10 MB says nothing about a slow link mid-path whose router queue overflows and drops silently. This is a genuinely separate bottleneck: a phone on a slow link has huge rwnd but tiny capacity; a slow app on fibre has the reverse. Neither number predicts the other, so no single mechanism can respect both:
The asymmetry that makes congestion control hard: rwnd is told to the sender by the party who knows it; cwnd cannot be — nobody knows the path’s capacity. It’s shared, changing, and (per IP) unreported, so it must be inferred, and the only evidence is loss:
Loss means congestion.
Not obviously true (loss could be corruption), but queue overflow dominates on wired networks — which is also why TCP historically did badly over WiFi, where loss often means interference and backing off is wrong. Mechanisms: AIMD (grow cwnd linearly, halve on loss — retreat fast, creep back) and slow start (double each RTT from a couple of segments until first loss; it starts small, not slow). Their reason for existing: the 1986 NSFNET congestion collapse, where senders retransmitting into a saturated network collapsed throughput while the network stayed maximally busy.
See also
- Every guarantee above IP is manufactured by state at the two endpoints — both windows are endpoint-held state;
cwndmust be inferred precisely because IP (the network) reports nothing - A cumulative ACK N means “I have everything through N” — the same ACK stream carries the advertised
rwnd, so flow control rides free on acknowledgment - IP promises nothing — it is best-effort and never reports delivery outcome — “loss means congestion” is inference forced by IP’s silence: the network won’t tell you it’s overloaded
- Reliability means preventing faults from causing failures — congestion control is what stops retransmissions (a reliability mechanism) from cascading into network-wide failure
References
Questions
flashcards/software-engineering/networking
What does flow control protect, versus congestion control?::Flow control protects the receiver’s buffer (limit rwnd); congestion control protects the network path (limit cwnd)
Why must cwnd be inferred while rwnd is simply told?::The receiver knows and advertises its free buffer (rwnd), but nobody knows the shared, changing path capacity — and IP reports nothing — so cwnd is inferred from loss
Data in flight is bounded by min(rwnd, cwnd) — the tighter of the receiver’s and the network’s limits.
What does AIMD do on success versus loss, and why the asymmetry?::Grows cwnd linearly while nothing is lost, halves it on loss — overshoot hurts everyone at the bottleneck, so retreat fast and creep back