Do we need this level of detail?

I'm not sure about the algorithm. I think what happens is cwnd tells us how long before we expect a response. If we send data for more than cwnd without a response that indicates congestion. So what we're doing is increasing the amount we either reach or desired rate as set by the ssthresh, or we don't get a response, indicating congestion.

If so, is this right?

If not, then we need to rewrite anyway, because it doesnt make sense.

I found these useful:

• https://www.researchgate.net/figure/TCP-Slow-Start-and-Congestion-Avoidance-phase_fig1_332436793
• https://www.noction.com/blog/tcp-transmission-control-protocol-congestion-control

So during slow start:

• cwnd grows exponentially for each ACK
• When we hit ssthresh, congestion avoidance begins and cwnd grows linearly instead
• If there's packet loss (timeout or duplicates), ssthresh set to cwnd/2 and cwnd is reset
• slow start begins again

I think we can say something like this for the:

Apparently cubic is the default on a lot of systems since quite a while:

https://en.wikipedia.org/wiki/CUBIC_TCP

The problem is that the last sentence might be true, but it isn't helpful without all that information you provided above. Reading it as it is written you'd jump forward to the section on collision avoidance and assume that this text was wrong - because the following sections says that slow start is one of the tools you can use for congestion avoidance.

I think the problem is ... the whole thing. We're trying to spot fix when we need to start again. Slow start is part of either "the" or "A" congestion control algorithm. It really doesn't make sense unless we talk about the whole thing.

Perhaps we can structure this https://www.researchgate.net/figure/TCP-Slow-Start-and-Congestion-Avoidance-phase_fig1_332436793 and even add a similar diagram?

Do you think we might do something like this for the whole thing (i.e. replacing all the current text).
I mean really, how much of our audience needs to know more than this. Though the diagram would be nice.

TCP slow start is the first phase of many common {{glossary('TCP')}} congestion control algorithms.
It is designed to rapidly determine a network's capacity at the beginning of a connection, after a period of inactivity, or after detecting congestion, without itself causing congestion.

When using the algorithm, data rates are set using a congestion window (cwnd).
This window defines the upper limit on how much data can be sent in which acknowledgment (ACK) packets are still being reliably received.
If responses are not received with the current cwnd, then the network is considered to be congested.

In the slow start phase the cwnd is initially set to a low value and increased exponentially until a configured slow start threshold (ssthresh) is reached.
In the second phase, "congestion avoidance", the cwnd is ramped up linearly until congestion is detected.
If at any point congestion is detected, ssthresh is set to half the level at which congestion was detected and the slow start phase is restarted.

I like this direction, so yes, I think you're right to go with your first impression to remove the specifics cwnd etc. I like the intro sentences we have, though, so we could cut even more details to summarize with a variation of this:

{{glossary('TCP')}} slow start is an algorithm that detects the available network bandwidth for packet transfer and balances transmission speed according to the network's capabilities.
The slow start mechanism prevents congestion when a network's capabilities are initially unknown, and slowly increases the volume of information sent until the network's maximum capacity is found.

Network congestion occurs when more data is sent than the network is able to transfer between nodes at any given time.
When a network is congested, machines on the network experience degraded quality of service, including slow data transfers and dropped packets.
The slow start algorithm is a major part of TCP congestion control, as it ensures that new connections progressively increase data rates and that those rates don't overwhelm other nodes.

Some points to keep, IMO:

• "prevents congestion when a network's capabilities are initially unknown"
• "slowly increases the volume of information sent until the network's maximum capacity is found"
• congestion effects are "degraded quality of service, including slow data transfers and dropped packets"
• "slow start algorithm is a major part of TCP congestion control,"

@bsmth I think this makes sense, or at least is a better baseline for future discussion. I have merged this in verbatim and if I have further comments I'll add them.

This looks great now, Hamish. I think the back and forth was worth it for the results 👏🏻

Phew. Thanks for your patience.