Agreed, although I see a lot of this work as input to future versions of a protocol (QUIC or otherwise) that could allocate bits like QUICv1 did for the spin bit. The IPPM document essentially is a zoo of the various ways bits could be used for spin-bit style uses, such that future protocols and protocol versions could decide to implement one. In the meantime, I agree that the application to QUICv1 is more for experimental demonstrations than something you could have arbitrary middle boxes assume support for by default.
Tommy > On May 11, 2023, at 4:50 PM, Christian Huitema <huit...@huitema.net> wrote: > > > > On 5/11/2023 3:53 PM, Martin Thomson wrote: >> On Fri, May 12, 2023, at 01:21, Christian Huitema wrote: >>> Igor is right. In fact, the "square bit" that he describes is >>> implemented in Picoquic, under an option negotiation. >> It's not about whether endpoints could negotiate something that allows the >> use of these bits, it's about whether measurement systems are able to rely >> on that happening. >> It's quite possible that some extension could be negotiated that creates a >> signal in these bits. OK, so now you can deploy a measurement system that >> relies on that. Most bits will be random, but some proportion will have the >> signal. Enter statistical methods and you can recover some signal. >> Well...almost. This works until someone else decides to negotiate the use >> of those bits for carrying a different signal. Now you have a harder >> statistical problem to solve. Maybe you can boost your adaptation by >> observing connection initiation and looking for clients that advertise a >> certain version and set of transport parameters. >>> The probability that these measurement features be turned on by default >>> is extremely low, and the draft should be very explicit about it, >>> including a description of the privacy/measurement tradeoff. >> This is really the point of all this. > > That. Which means that at a minimum the IPPM draft should discuss the issue: > needle of measurement in a haystack of noise. It might work sometimes, such > as when debugging the end to end path between two well known IP addresses, or > when knowing the specific connection IDs used by the system participating in > the measurement. It might perhaps work if the connection last long enough to > distinguish the IPPM pattern from a random pattern. But as you say, the > general case is almost impossible. > > -- Christian Huitema >