How TCP congestion control works
A sender can't see the network, so it probes: double fast, then creep up, and halve the moment packets get dropped. Each step here is one round trip.
Tweak it
Break it
Normal run: TCP Reno-style. Slow start, then +1 per round trip, and halve on loss.
What to notice
- Slow start doubles the window every round trip, which is exponential, until it reaches the threshold. Then congestion avoidance adds one per round trip.
- On loss the window is cut in half. The classic sawtooth is TCP probing for the capacity and backing off.
- Raise the random loss slider: the connection never reaches capacity, because every loss halves the window.
- Break it: with no backoff the link is overloaded, so goodput (useful data delivered) falls. Senders backing off is what stops the network from collapsing.
- Simplified model: Reno-style, one step per round trip, and the collapse curve in Break it is illustrative, not measured.