[rayiner]

> Instead of carelessly flinging packets into the ether like an savage, you had a deterministic network of pipes

I love this. Ethernet is such shit. What do you mean the only way to handle a high speed to lower speed link transition is to just drop a bunch of packets? Or sending PAUSE frames which works so poorly everyone disables flow control.

[packetlost]

Wait, are you serious? This is how it works?

[rayiner]

Yes: https://fasterdata.es.net/performance-testing/troubleshootin.... A simplistic TCP server will blast packets on the link as fast as it can, up to the size of the TCP receive window. At that point it’ll stop transmitting and wait for an ACK from the client before sending another window’s worth of packets.

To handle a speed transition without dropping packets, the switch or router at the congestion point needs to be able to buffer the whole receive window. It can hold the packets and then dribble them out over the lower speed link. The server won’t send more packets until the client consumes the window and sends an ACK.

But in practice the receive window for an Internet scale link (say 1 gigabit at 20 ms latency) is several megabytes. If the receive window was smaller than that, the server would spend too much time waiting for ACKs to be able to saturate the link. It’s impractical to have several MB of buffer in front of every speed transition.

Instead what happens is that some switch or router buffer will overflow and drop packets. The packet loss will cause the receive window, and transfer rate, to collapse. The server will then send packets with a small window so it goes through. Then the window will slowly grow until there’s packet loss again. Rinse and repeat. That’s what causes the saw-tooth pattern you see on the linked page.

[wat10000]

This is how old-school TCP figures out how fast it can send data, regardless of the underlying transport. It ramps up the speed until it starts seeing packet loss, then backs off. It will try increasing speed again after a bit, in case there's now more capacity, and back off again if there's loss.

[bombcar]

You can achieve a bit of performance here by tuning it so it will never exceed the true speed of the link - which is only really useful when you know what that is and can guarantee it.

[cylemons]

TCP still works this way?

[Spooky23]

I experienced this with a VDI project when we mistakenly got 25Gb links delivered to the hosts.

We were expecting to get some sort of unbelievably fast internet experience, but it was awful as the internet gateway was 1 Gb or something similar.

[EvanAnderson]

Heh heh. If that shocks you, search engine for "bufferbloat" and prepare to be horrified.