Why a download starts slow
Start a large download and watch the speed. It does not begin at full speed. It climbs: slowly for an instant, then quickly, then it wobbles around some level for the rest of the download.
That behaviour is not your connection warming up. It is a deliberate search, carried out by a part of TCP called congestion control.
The problem: a network you cannot see
When your computer sends data to a server on another continent, the packets cross a dozen networks owned by different companies. Each link has a different capacity, and each is shared with a changing crowd of other users.
The sender has no way to ask "how fast can I go?". Nobody along the path knows the whole answer. If it sends too slowly, the transfer wastes time. If it sends too fast, the busiest link overflows, packets are dropped, and they have to be sent again, which makes the congestion worse for everyone.
So TCP finds the right speed by experiment.
The window
The sender keeps a number called the congestion window: how many packets it is allowed to have in flight, sent but not yet acknowledged. A bigger window means a faster transfer.
The whole algorithm is about how to adjust that number using the only signal available: did my packets get through?
Start. A download begins. The sender has no idea how much this network can carry, so it starts by sending just 1 packet and waits for the reply.
- Round trip
- 0
- Window (packets)
- 1
- Phase
- slow start
- Losses
- 0
Phase 1: slow start
Begin with a tiny window. Each time a full window of packets is acknowledged, double it.
1, 2, 4, 8, 16, 32, 64.
It is called slow start because it starts small, not because it grows slowly. Doubling reaches high speeds within a handful of round trips. This is the brief ramp-up you see at the start of a download.
The signal: a lost packet
Sooner or later the window exceeds what the path can carry. A router somewhere runs out of room and drops a packet. The sender notices that an acknowledgement is missing.
TCP treats that loss as a message: "you have gone too far."
Phase 2: back off, then probe gently
Cut the window in half
A large, immediate reduction relieves the congestion.
Remember roughly where the limit was
The capacity is somewhere near where the loss happened.
Grow by just one packet per round trip
No more doubling. Creep upwards.
On the next loss, halve again
And repeat, for the rest of the transfer.
The result is the sawtooth in the chart: a slow climb, a sharp drop, a slow climb. The window is never exactly right. It hovers just below the limit, repeatedly reaching up to test whether more room has appeared.
Example
Why keep testing?
Capacity is not fixed. Someone else on your network finishes a video call and bandwidth frees up. The gentle climb finds it within seconds. If someone starts a call, the next loss makes the sender give some back.
The sawtooth is the protocol continuously re-measuring a network it cannot see.
Why halve, but only add one?
Being cautious on the way up and drastic on the way down is deliberate. It is called additive increase, multiplicative decrease.
Picture two downloads sharing one link, one with a much bigger window than the other. When the link overflows, both halve. Half of a big number is a big cut. Half of a small number is a small cut. Then both add one per round trip. Each cycle, the gap between them shrinks.
Repeat that and the two drift towards equal shares, with nobody coordinating them. The same rule, run independently by billions of devices, is a large part of why the internet does not collapse under its own traffic.
What you notice from outside
- Short transfers never reach full speed. A small web page finishes while still in slow start. This is one reason opening a fresh connection for each small request is wasteful, and why connections are reused.
- A lossy link is slow even if it is fast. On poor Wi-Fi, packets are lost to interference, not congestion. Classic TCP cannot tell the difference and halves anyway.
- Latency matters. The window grows once per round trip. A distant server means longer round trips and a slower climb.
Key takeaways
- A sender cannot know the capacity of the path, so TCP discovers it by experiment.
- The congestion window limits how many packets may be in flight at once.
- Slow start doubles the window every round trip until a packet is lost.
- After a loss the window is halved, then increased by one per round trip.
- The resulting sawtooth keeps the sender just below the limit and shares the link fairly.