Speaking out loud in case anyone sees this, but regarding jumbo frames:
why aren't they the default since, I don't know, fiber ISP rollouts? Outside of a slight increase in latency and maybe loss of value with small webpages/fragments, surely ISPs can handle it and you could upgrade fiber ONTs to easily handle 2.5, 5, 10 GigE or fiber directly to your internal network. There's almost no excuse. Other than I guess resource starvation of the end device processing data. Am I not understanding something?
To summarize for the uninitiated, I learned today that IPv6 supports packets of up to 4GB! Why you would ever need a packet within even an order of magnitude of that size is beyond my ken. (Also, I don't know what transit media even support packet sizes anywhere close to that size).
The only place I could see the layman ever getting close to using this would be loopback. But at that point, you have IPC.
Also I assume it’s not a design goal to support 4GB but it is just a result of a power of 2. They are basically just doubling the size from 16 to 32 bits. Trimming it to 30 bit to cap it at something "reasonable" 1GB or something is more hassle then it’s worth.
I just cannot imagine the error correction on those packets. I mean if anything goes wrong that's a lot to toss away.
> I just cannot imagine the error correction on those packets. I mean if anything goes wrong that's a lot to toss away.
I think the original jumbogram work occurred back when alot of people were still convinced end-to-end IP alone wasn't suitable for many uses. This was the era when, e.g., people were still pitching ATM or exploring bespoke solutions for linking research centers. So basically jumbograms were presumed to be delivered atop another long-haul, multi-link protocol that would provide reliability. Jumbograms were a way to improve the throughput and latency between IP peers with minimal but high-impact changes to application and kernel network stacks (presuming IPv6 adoption), placing most of the complexity in the network transit layers. What we ended up with, ironically, was the reverse, everything IP and below largely ossified from software perspective, some minor evolution at the TCP layer, but most progress happening above TCP/UDP, and alot of work invested in local hardware link layers that remains opaque to IP or even the ethernet software stack, effectively placing a ceiling on dividends.
I hadn't thought of that, that would be a potential use case for sure. And I agree that literal 4GB packets probably aren't what's in mind here, but I think it is an oddity that a confluence of other factors has it allowed by the RFC.
https://www.rfc-editor.org/info/rfc1149/ has more specific information in regard to the benefits of avian carriers. Particularly "because of the 3D ether space available to the carriers, in contrast to the 1D ether used by IEEE802.3."
Thank you riffic for sharing this, this reminds me of the SIGBOVIK papers :)
Is there really that much room for improvement by changing packet size? You're avoiding a little bit of overhead for the 40 byte packet header and you're also increasing the risk of that packet failing CRC and getting discarded which is a lot more expensive.
I mean over the standard 1500 byte packet, for sure. I'd assume that you'd probably find an optimal spot at 1 to 8Mb on super fast networks. Also networks at these speeds tend to have very low retry rates. It's the buffer sizes that start to get crazy though.
Sigh. IPv6 again, the tale of bad design and stupidity. IPv6 is forever stuck with 1462 MTU now because they neglected the PMTU discovery.
For larger packets to work, the MTU size signal has to be baked into _each_ packet. ICMP on truncation is NOT a good design.
Of course, there's a belated RFC to fix it: https://www.rfc-editor.org/info/rfc9268/ But it'll never get implemented and deployed, as we're deep in the "ossification" stage for IPv6.
what you you mean they neglected discovery? Everyone knows that IPv6 stops working if you block ICMP because it blocks MTU discovery, ergo everyone knows it relies on MTU discovery.
No standard equipment can do 64K+ frames, so you are not going to be connecting to global internet soon.
Plus, why are you doing this at all?
If you are super-concerned with the overhead, you don't want IP stack anyway, with all its layers and tunables - you want something like RDMA which can work very close to metal.
And if you are OK with overhead, then why go with huge frames? Go with small ones and reassemble, your life would be much easier. (Normally reassembling is a pain because of possibility of lost or reordered packets, but if you are willing to go with such huge packet size, it must not be a concern with you)
(this reminds of the opposite direction: people stuffing IPv6 into sensor networks.. your frames are 32/64 _bytes_ and you want IPv6 headers in there???)
It's more convenient if there's a standard layer available. It means you can introduce new things at L1/L2 and they'll still work with all existing software, and also you can introduce new things at L4+ and they'll work with any L1/2 tech, so long as you make your new thing map to L3 in the middle.
> (this reminds of the opposite direction: people stuffing IPv6 into sensor networks.. your frames are 32/64 _bytes_ and you want IPv6 headers in there???)
You mean like in 6LoWPAN, which uses header compression to get them down to 2 bytes? While still working with regular TCP- or UDP-using software?
That was my read on it. I had no idea this was part of the protocol (I saw it while reading the IPv6 Wikipedia article). Like I said in my sibling comment, maybe, maybe, you could use this for loopback, but I can't think of a single use case that actually justifies its existence.
why aren't they the default since, I don't know, fiber ISP rollouts? Outside of a slight increase in latency and maybe loss of value with small webpages/fragments, surely ISPs can handle it and you could upgrade fiber ONTs to easily handle 2.5, 5, 10 GigE or fiber directly to your internal network. There's almost no excuse. Other than I guess resource starvation of the end device processing data. Am I not understanding something?
The only place I could see the layman ever getting close to using this would be loopback. But at that point, you have IPC.
I mean inter-container communication on a host?
Also I assume it’s not a design goal to support 4GB but it is just a result of a power of 2. They are basically just doubling the size from 16 to 32 bits. Trimming it to 30 bit to cap it at something "reasonable" 1GB or something is more hassle then it’s worth.
I just cannot imagine the error correction on those packets. I mean if anything goes wrong that's a lot to toss away.
I think the original jumbogram work occurred back when alot of people were still convinced end-to-end IP alone wasn't suitable for many uses. This was the era when, e.g., people were still pitching ATM or exploring bespoke solutions for linking research centers. So basically jumbograms were presumed to be delivered atop another long-haul, multi-link protocol that would provide reliability. Jumbograms were a way to improve the throughput and latency between IP peers with minimal but high-impact changes to application and kernel network stacks (presuming IPv6 adoption), placing most of the complexity in the network transit layers. What we ended up with, ironically, was the reverse, everything IP and below largely ossified from software perspective, some minor evolution at the TCP layer, but most progress happening above TCP/UDP, and alot of work invested in local hardware link layers that remains opaque to IP or even the ethernet software stack, effectively placing a ceiling on dividends.
A 4 GB microSD card is well within the carrying capacity of a pigeon.
https://www.rfc-editor.org/rfc/rfc6214.html
Thank you riffic for sharing this, this reminds me of the SIGBOVIK papers :)
For larger packets to work, the MTU size signal has to be baked into _each_ packet. ICMP on truncation is NOT a good design.
Of course, there's a belated RFC to fix it: https://www.rfc-editor.org/info/rfc9268/ But it'll never get implemented and deployed, as we're deep in the "ossification" stage for IPv6.
No standard equipment can do 64K+ frames, so you are not going to be connecting to global internet soon.
Plus, why are you doing this at all?
If you are super-concerned with the overhead, you don't want IP stack anyway, with all its layers and tunables - you want something like RDMA which can work very close to metal.
And if you are OK with overhead, then why go with huge frames? Go with small ones and reassemble, your life would be much easier. (Normally reassembling is a pain because of possibility of lost or reordered packets, but if you are willing to go with such huge packet size, it must not be a concern with you)
(this reminds of the opposite direction: people stuffing IPv6 into sensor networks.. your frames are 32/64 _bytes_ and you want IPv6 headers in there???)
AKA: https://en.wikipedia.org/wiki/Hourglass_model
> (this reminds of the opposite direction: people stuffing IPv6 into sensor networks.. your frames are 32/64 _bytes_ and you want IPv6 headers in there???)
You mean like in 6LoWPAN, which uses header compression to get them down to 2 bytes? While still working with regular TCP- or UDP-using software?