Re: [quicwg/base-drafts] QUIC lacks on-path exposure of packet loss (#3189)

[Alexandre Ferrieux <[e-mail]>]

@britram:

Re "okay, let's stick with the metric we know": there is a bit more do be said. In real-world, heterogeneous/not-too-modern networks, we are essentially stuck in a pre-ECN world; there, even the smartest of congestion control algorithms must end up interpreting some loss patterns as congestion, and sometimes, wrongly.

Re privacy analysis: indeed this is what we are explicitly asking for. But using the Spin Bit as a reference to infer the amount of work needed is not really appropriate I think, as the very nature of the exposed signal is entirely different.

Re using the bits differently: you'll surely remember that, as an early admirer of the sheer beauty of the VEC, I certainly have no preconceptions against the principle. However:

1. As Marcus' research showed, it turns out that observer-side denoising does the job reasonably well, with a definition of "reasonably" that led to the current consensus, a one-bit Spin Bit.

2. Firsthand experience with debugging in the aforementioned real-life networks in several countries and varied multi-stakeholder scenarii tells that loss is *much* harder to locate than latency (see this Issue's introductory text for details). 

3. No, dropping one of the bits is not an option: the key feature of the method is *dichotomy*, meaning we need to somehow convey measurements about either side of the observer. And any ECN-based "trick" to emulate Q and carry only L is, again, predicated on the availability of ECN-capable layer 3 on the uninterrupted path.

As a result, I believe that the Spin+Q+L combination is, given a 3-bit budget, the best compromise to support network troubleshooting.

-- 
You are receiving this because you are subscribed to this thread.
Reply to this email directly or view it on GitHub:
https://github.com/quicwg/base-drafts/issues/3189#issuecomment-550494396