Notes taker: Marwan Fayed
*Robert Richter (10 mins)
Mike Blanche: Thank you. What's the list of ASes on the right side?
Robert: Most frequent hop that I mapped. Tried to answer "how much does Starlink move and which ASNs involved"
Mike: Hop 9/10/11, AS5680 - they're a website provider?
Robert: Yes. They probably have some service that is used.
Mike: You are tracerouting to k root?
Robert: Yes
Lorenzo: I think this is not representative of traffic. If you see a traceroute then need to traceroute from elsehwere. Also ping measurements, with los of 5% you can't get reliable measurement; you probably want to discount from measurements. If there was really 5% packet loss then TCP would fall apart.
Mirja: Disagree. This is just ping loss.
Tianji: From UVic they do testing. They do measurements from one side of country to another, but find packets travel in the opposite direction. That would give the long latency observed.
Mirja: I think conclusion is you need to learn more about how Starlink works.
Robert: Yes, definitely.
Mike Blanche: Thanks. How are you determining cache hits and misses?
Nitinder: x-cache status in the response
Mike: Different services are distributed across different locations, e.g. video caching. It's interesitng that CDN platforms will have more content in places where it is accessed more often, and not in others. So if you're routing to different countries (e.g. with different language) then you are more likely to miss.
Nitinder: Thank you.
Marwan: Just a comment about putting caches in space -- it seems a natural thing to do, but as net and sys people it's easy to overlook the challenges of server authentication, for example what it means to have private keys flying over planet, or keyless signatures, etc.
Nitinder: Yes, of course. This was pitched as a HotNets idea, and we can push the envelope to PQC, etc. All hard problems that deserve attention.
Mirja: Sounds like a new research question!
(no questions)
Mirja: Did you reach out to the implementers and get feedback?
Jonas: We did reach out with some bugs and tried to get them fixed but we did not further talk about this instant ack deployment.
Eric Kinnear: Thanks this is cool work. When you talked about types of packet lost, did that cover all the permutations?
Jonas: We looked at the most important ones. IACK also has effect on later packet exchanges, but the initial ones are most important. The effects on later transactions is less.
Mirje: It's really important to compare different stacks. Also from chat, complaints that some stacks were not evaluated; also that data is hard to decouple from the protocol and always something needs to be sent, so it would be good to figure out what that looks like; also from chat implementers are encouraging you to get in touch directly.
Roland: Thank you.
Mirja: from chat, qlog is very useful
Marcel: Packet traces are on github. We did not enable qlog.
Mirja: Yes, a suggestion for the future.
Marwan: Sometimes tc can be a source of measurement peculiarities; did you take that into account?
Marcel: Yes, that's why we had the fibre tap so we could observe without noise from the client.
Marwan: Ah, right, thank you.
Chris Box: I'd like to check what you think is the ideal pacing strategy? I tend to agree if you're talking about wireless network, then single packet with pacing between them. But if other medium in which when you have access it's better to batch send, it may be that perfect pacing is less than ideal and burst send is better.
Marcel: Agree. We discussed a lot, and found it's not so clear. It cannot be answered easily. For example, the opposite is "what is a burst?"
Chris: Ok, so it's an open question then.
Marcel: Yes.
Mirja: Do you plan to keep working on the kernel patch.
Marcel: We plan to continue to look at pacing, but no plans for the kernel patch. As I said ours was a modified years old Google patch. It's not clear how to proceed with it.
Mirja: But if others interested they can reach out to you.
Marcel: Yes, definitely.