What it is
A TCP sender does not know the capacity of the path. It raises its congestion window until a packet is lost, then backs off: Reno adds one packet per round trip and halves on loss, CUBIC grows along a cubic curve in elapsed time and backs off to 70 %. With a drop-tail buffer the window oscillates in a sawtooth, the queue fills and empties, and several flows end up sharing the link. This simulator runs 1–4 flows over one bottleneck at packet level.
When to use it
Use it to see why a buffer of one bandwidth-delay product keeps a link busy, why a much deeper buffer only adds delay (bufferbloat), how throughput on a lossy path follows loss and round trip time rather than link speed, why a short and a long flow share unevenly, and what a single lost packet or a burst of ten costs. Drop packets yourself, halve the capacity mid-run, rewind, and compare Reno, CUBIC and AIMD(a, b) on identical conditions.
Gotchas & limits
This is a teaching model of congestion control on one drop-tail bottleneck: BBR is deliberately not modelled, loss recovery is SACK-like (the sender knows exactly which packets are missing) rather than real SACK blocks, and delayed ACKs, pacing, HyStart, ECN and tail-loss probes are left out. Constants such as the initial window of 10, the 200 ms RTO floor and CUBIC’s 0.4 and 0.7 are recalled from the RFCs, not re-checked. Results are checked against closed forms (the Mathis formula, AIMD scaling, CUBIC’s K), not against a real network or kernel.