Hacker News new | past | comments | ask | show | jobs | submit login

> Now, if only we could get jumbogram more widely deployed. Thats older than V6 is and still struggling to break the 1500 byte MTU limit.

Really, we barely hit 1500. Look at mss for popular websites, most people drop from 1500, because 1500 has problems in enough places. Does http/3 even send 1500 byte packets ever?

One major problem is most servers (Linux all versions, I think, FreeBSD before about 2000 and after something like 2019) always send the interface mss with a syn+ack. You get meaningfully better results by sending the lesser of the interface mss and the received mss; there are enough broken systems out there that don't communicate the real MTU to end systems[1], don't send enough ICMP needs frags, try to cover it up by manipulating mss in syns, but don't manipulate mss in syn+ack. Windows and iOS (and presumably mac) do a pretty good job of detecting pmtud blackholes, but it's often disabled on servers and I can't remember where Android is these days; I know it used to ship with the option compiled in but disabled and no way to enable.

Of course, packet sizing is actually a hard problem. Larger packets are good for faster links but bad for slower links.

[1] which is hard, because I don't know how you get windows to use MTU from dhcp; it doesn't request it, so it won't use it. This is a problem too.




> Does http/3 even send 1500 byte packets ever?

Iirc quic has a hard MTU cap at 1280 bytes.


Nope. QUIC requires the _minimum_ MTU of 1280: https://datatracker.ietf.org/doc/html/rfc9000#name-datagram-...

The maximum permitted size is 65527 (max_udp_payload_size).




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: