Mirja Kühlewind and Dave Plonka, chairs
Brian Trammell, scribe
Ranya Sharma (15 mins)
Tony Tauber: if there were concurrent users on the wifi link would you
get more utilization?
Ranya: we didn't look at concurrency, good next step, but we would
expect that the bottleneck would persist.
Greg White: understanding where the bottleneck is is useful to find
buffering delay for latency. Would be nice to see actual delays on
different segments of the network. Interesting to look into this over
time as new wifi generations roll out.
Ranya: good next steps, thanks.
Arturo Filastò, OONI: is a sample size of 52 sufficient for these
conclusions
Ranya: nope, want to expand beyond this sample which is concentrated in
Chicagoland.
Arturo: We have a measurement platform. It might be helpful.
Tanmay Rajore (15 mins)
Arturo Filastò: could future work also look into blocking resistance
related to the use of QUIC/MASQUE?
Tanmay: already looking into this, happy to talk to you more about this.
Arturo: Cloudflare Warp and Apple Private Relay use this.
Tanmay: Currently in development. Cloudflare uses Wireguard
Lorenzo Colitti: Looked at TLS over TCP only? Not QUIC? Obviously this
will work badly if there's UDP in the VPN. One attack, you can correlate
packets in and packets out if you're tapping both sides of a DC. Have
you seen any implementations using more than one VPN server to defend
against this?
Tanmay: ProtonVPN and NordVPN recently moved to two-hop.
Lorenzo: Something we're looking at is have the app contract with its
own server, and connect with that, on a per-app basis. When each app has
its own VPN server that raises the correlation bar.
Weitong Li (10 mins)
no questions
Johannes Zirngibl (10 mins)
Arturo Filastò: why do we need IPv6?
Mirja: take it offline please.
Dave Plonka: The opportunity is now that the tools exist to help
implement the more difficult Happy Eyeballs v3. very cool.
Tommy Pauly: curious to see whether the order of the tests influences
historical caching to help validate the caching logic. When you're
changing the value a lot in these tests, you're messing with that cache.
When you have different domains do they always map to different
addresses?
Johannes: all to the same prefix, one address per delay.
Tommy: (advertises HAPPY), how can we use this tooling to update the
numbers in the document, looking at those, to get recommendations based
on real world impact.
Johannes: that's your job, you have the browser data. :)
Vaibhav Bajpai: We have an HEv3 impl, we'd be happy to plug this in.
Thom Vaughan (10 mins)
Atruro: could you give me a reference for harmonic centrality? (offline,
link in chat: https://arxiv.org/abs/1308.2140)
Mirja: Do you provide the analysis somewhere?
Thom: Link to the S3 bucket in the slides.
Syed Suleman Ahmad (remote) (10 mins)
no questions
Tommy Pauly (10 mins)
Marco Munizaga: does this happen with every protocol version change?
Tommy: It's essentially the same set that updates every time. We'll end
up with two protocol stack eventually.
Momoka: For v4 vs v6 data, are we looking at dual stack or not?
Tommy: We're only looking at dual-stack connected clients here.
Vaibhav: need to show correlation without pie charts. I expect there are
a few variables hiding behing the scenes.
Tommy: This is what I can show. Client is offering the same initial in
all of these cases; won't vary per network where the client is. Not
trying to tease apart the confounding variables.
Allison Mankin: Potentially futile question, can you make anonymized
data available?
Tommy: would love to. can't offer that now. would be interesting to have
a
Mirja: maybe people could approach you (e.g., from maprg) about data
access?
Tommy: if there are things [you'd] want to see in future maprg
presentations, maybe that's a start