Wednesday 8 November 2023 09:30 UTC+1 --- Congress 2 Hilton Prague
Chairs: Dave Plonka, Mirja Kühlewind
Notes: Brian Trammell, Ryo Yanagida
(no comments)
Roland Bless (10 mins) slides
Michael Tüxen: you're suggesting to go multicore, is the TCP case
multicore for a single connection?
RB: no.
Alan Frindell: thanks and greetings from mvfst. our defaults are
probably not performance optimized. tuning quic is hard, but it looks
like we have some work to do.
RB: yes, we thought that might be contributing to the results here.
Geoff Huston: TCP line card segmentation offload has a dramatic
impact. QUIC doesn't do seg offload. In your measurements was line card
segmentation enabled?
RB: offload enabled. Impact not huge though.
GH: we've found this is interesting on high packet load. Interrupt
bandwidth.
RB: 100g experiments, more important
Theo Benson (10 mins) slides
Brian Trammell: I'd like to encourage the community to follow this
closely. Both of these are looking at very early implementations. It
would be interesting to see how much faster QUIC gets over time. There's
work from a while ago on getting TCP to 100M, to 1G, to 10G, and so on,
and comparing the performance improvements would be interesting.
Ike Kunze, Constantin Sander (15 mins) slides
Brian Trammell: Great work. Spin-bits over-estimation in your results
seems a little high. How do you get the ground truth? ping?
Ike: Compared to the RTT estimation from the QUIC stack itself.
Marten Seemann: Bit surprised to see the results here, we did a lot of
evaluation during hackathons, and we found the spin bit RTT estimates to
be more accurate than that.
Ike: We have a theory, looked at short H3 connections, where the
application delay might dominate.
Tara Tarakiyee: What does "the spin bit is used by these domains"
mean?
Ike: A domain has the spin bit enabled if it is used; if we see one
zero and one one then we try to evaluate it
Tara: Wouldn't "is used" mean "people are using it to measure RTT"
Ike: We're just looking at whether the remote domain supports
spinning.
Maynard Koch (10 mins) slides
Lorenzo Colitti: We don't ship or run forwarders. Can you go back to
the invalid source address arriving at the host? How would the client
actually get the reply from the unexpected source? Android would just
reject it, because it's using a connected socket.
Maynard: the client should reject it yes.
Lorenzo: Why are people doing this if it's rejected? Just mistakes?
Maynard: Seems to be due to implementation errors.
Lorenzo: 600k of these that don't do anything? Wow.
Mirja: Let's take this offline.
Sudheesh Singanamalla (remote) (15 mins) slides
Maynard Koch: On slide 6, surprised by why shadowserver sees 2.5M and
you see a lot more, this is probably due to the scanning setup. We also
see about 10M resolvers but we filter out some resolvers, after
filtering we see about 2.5M
Geoff Huston: You borrowed a slide from APNIC, but that is not a
measure of resolvers, but rather, end-users and resolvers end-users use.
Are there really any end-users in those open resolvers? Might be looking
at random noise, these might be things that just pop up on port 53. Try
correlating these with known lists of known resolvers with known
clients.
Sudheesh: Agreed, that is a concern we have. Do they even have clients
or not is important — if they to, there is a risk.
Weitong Li (remote) (15 mins) slides
(no questions)
Maxime Piraux (10 mins) slides
Simon Leinen: first reaction to slide 18 to first hop RTT change in
the peer network is that the return path must be different. Assume v6
traffic has an asymmetric return path. Warrants some thought, because
aggregation and therefore TE possibilities are different. Topology is
not very different, may have been ten years ago. But there are
differences in routing practices. Your oberservations are a good
argument for multipath.
Tommy Pauly: thanks for doing this work, interesting to see another
way to look at these numbers. Our metrics for v4/v6 look very close. v6
is a little faster, but correlates highly with other things that make
networks better, like more modern stacks. You mentioned HE3, the use of
priority allows us to maybe indicate the quirks of which works better or
not. Priority in the algorithm still in flux, please write to the v6ops
list.
Mirja: Tommy, refresh your measurements in MAPRG next time please.
Eric Kinnear: Interesting things you did was the comparison between
the v4/v6 on a same-path. What we see in our result is that v6 wins the
HE race. We still stick with ipv6 unless the v4 latency is tens of
milliseconds lower, for non-latency reasons.
Johannes Zirngibl (10 mins) slides
(No questions)
Yasir Zaki (remote) (10 mins) slides
(No questions)
-meeting end-