Never seen used in the practice though. One of problems AFAIK - most switch default to dropping jumbo frames. You can enable but if you'll left one unconfigured for some reason you will get hard to diagnose problem. The same with end hosts - they (inside a given L2 domain) should have the same MTU and if you'll left a box with the default MTU you got a problem.
To make jumbo frames easier to use switches should forward them by default and hosts should accept frames large than MTA (but send MTU sized ones).
Jumbo frames are used constantly in practice, just not often on the internet. They're common in all sorts of networks, especially enterprise and storage.
If your switch is configured on defaults and not running a for purpose config, you probably don't need jumbo frames.
I've been terrified to enable jumbo frames because every guide I've read has always had "you have to make sure that every single machine you ever communicate with has to also have jumbo frames on or you will cause a black hole that sucks the whole earth into it!!" kind of disclaimers when I just want to lower the overhead in copying files from my NAS that has some pathetic Atom CPU in it.
If that's the case it's a shame we didn't take the chance with IPv6 to push to higher default MTUs since IPv6 already relies to a much higher degree on PMTUD since it lacks fragmentation.
the internet has a whole bunch of non ethernet stuff, a lot of which has different frame sizes. Its totally possible that backhaul is running jumbo frames, or something like it, but you'd never really know that.
Conversely ADSL has odd frame sizes(inherited from ATM; 48 bytes if I remember correctly), but you don't see that because its hidden from you. Cable has a frame sizes ranging from ~500 up to 2000 bytes. Again, hidden from you.
One of the joys of TCP/IP is that different frame sizes are handled for you. Sure it might be beneficial to have a frame size that marries up with packet size, it might not. you don't really know, because the internets.
The bottom line is that if you ever send an IP frame larger than 1500 bytes outside of your own network, it's most likely that it will never reach its destination. Especially for IPv6.
I wouldn't call them "frame sizes" for ATM or DSL. It's a bit of a transparent fragmentation into "cells" onto the "transport" layer. It could carry arbitrary "frame" sizes on that
It being a thing on "the internet" would require it being a thing on a sizable majority of the internet. You'd need to get large network operators, peering points, user-facing ISPs, and even users themselves on board to change their setups.
And Path MTU discovery is still sufficiently unreliable as to make it incredibly painful to have partial large-MTU networks.
And if you do any of this with standard home customers, a hellscape torrent of user complaints is going to rain down on your support contacts. Which costs money. More money than is lost by the higher cost of routing smaller packets.
So, no, jumbo frames could not be a thing on the internet. There's a reason it's called fossilization. There is no technical reason precluding changing this, it's just frozen into way too many places to be changed.
This the same argument why IPv6 can't be a thing on the internet; I agree that the book isn't closed entirely on that yet, but very significant progress has been made.
Except disruptions (i.e. worse service than IPv4 only) from rolling out IPv6 are the exception while disruptions from rolling out jumbo frames are absolutely the norm.
I thought frame size negotiation was an optional part of LLDP now. I suspect it fits into the broad category of things that could be well-defined if there was broad adoption of a poorly adopted standard.
