Scenarios

"Slow start" doubles the window every round trip. It is slow to begin and then explodes, and on a shallow buffer it overshoots the pipe badly before the first loss tells it to stop.

Reno adds one segment per round trip, so the time it needs to fill a pipe grows with the RTT. On 100 Mbps at 100 ms the window has to reach about 860 segments, and one-per-RTT will not get there. CUBIC grows as a cubic function of time since the last loss, so its recovery does not depend on how far away the other end is. This is why Linux changed its default.

A loss-based sender fills whatever buffer it is given — that is what "loss based" means. Give it eight BDPs of buffer and it will put eight BDPs of data in the queue, adding hundreds of milliseconds of latency without losing a single packet. This is why a video call degrades when someone in the house starts a download.

Chiu and Jain proved in 1989 that additive increase with multiplicative decrease converges to an equal share from any starting point. The second flow here arrives 20 seconds late, so it starts from as unequal a position as there is.

Window growth is per round trip, so a flow with a quarter of the RTT increases four times as often and takes a correspondingly larger share. The two flows here are identical in every other respect.

BBR v1 does not treat loss as a congestion signal. It measures the bottleneck bandwidth and the propagation delay and paces at that rate. Sharing a deep buffer with a loss-based flow, the two are optimising for different things — and how the capacity ends up divided is an actively contested question in networking, not a settled one.

Vegas watched queueing delay and backed off before anything was dropped — in 1995, more than twenty years before BBR. It works: on its own it fills the link and leaves the queue almost empty. But it shares a bottleneck with a loss-based flow that only backs off when it loses a packet, and the capacity Vegas politely declines is simply taken by the flow next to it.

Without selective acknowledgement the receiver can only say "still waiting", so the sender fills one hole per round trip. After a burst loss on a long fat path that is hundreds of round trips. With SACK the sender knows exactly which segments are missing and fills them all at once.