Hacker News Reader

Comment by toast0 | Hacker News Reader

You can certainly chunk things and interleave/multiplex interactive and bulk data.

But it's easy to end up with low throughput, because you limit how much unacked bulk data you send, because you don't want to overqueue bulk data and not be able to immediately write interactive messages.

And it's easy to end up with high latency, because despite the limits above, you queued too much and interactive messages need to wait. Or you sent bulk data and there was a burst of packet loss and new data can't be received until the missing packets are resent and received.

> people did not start writing for TCP/IP yesterday.

And yet you find resuming uploads over http to be too hard?

View on HN

>But it's easy to end up with low throughput,

>And it's easy to end up with high latency,

And yet the OS is doing just that when you open two TCP streams. It packs them into ethernet frames and keeps track of the acks. Only the OS does not know the context, so it prioritises both streams equally. (Unless some evil magic is involved.)

>And yet you find resuming uploads over http to be too hard?

Me? No. Xmpp developers? Apparently yes, as it's neither in any of the XEPs nor in any of the clients.

> Only the OS does not know the context, so it prioritises both streams equally.

Of course there's prioritization options on socket APIs since forever (DSCP), but usage is limited, because it requires coordination and coordination is hard.

Even with equal prioritization, when you overuse on the media connection and that results in packet loss and congestion window reduction, chat messages don't have to wait for all the dropped packets to be resent over multiple round trip times. If you're lucky the chat connection didn't see any of the loss, and even if you're not, you've likely got a smaller queue on your chat connection than your media connection.

This also doesn't prohibit client driven prioritization --- when connecting, get caught up on chat before processing media. Maybe pause media while sending chat messages or if server acks are slow. Server driven prioritization is hard, because probably your http media servers don't have information on the chat connection --- but if a client stops reading from the media connection, the server will get the message eventually.

>Even with equal prioritization, when you overuse on the media connection and that results in packet loss and congestion window reduction, chat messages don't have to wait for all the dropped packets to be resent over multiple round trip times.

Well, again, it's merely an implementation detail. You could just open a second TCP connection using the in-band IM protocol, not some external HTTP. Or make your IM protocol UDP based.

But really, that's what quic was designed for, which is also not new.

Reply on news.ycombinator.com