Re slide 8 - why did idle latency drop in June 2024? That is when Starlink implemented AQM on the home gateway.
", "time": "2025-07-25T07:43:28Z"}, {"author": "Jason Livingood", "text": "It is not a Q
", "time": "2025-07-25T07:43:55Z"}, {"author": "Jason Livingood", "text": "(slide 8 asked the Q - I tried to answer it)
", "time": "2025-07-25T07:44:19Z"}, {"author": "Dave Plonka", "text": "ah, tnx. hard to track the content while chairing (for me) :)
", "time": "2025-07-25T07:44:37Z"}, {"author": "Jason Livingood", "text": ";-)
", "time": "2025-07-25T07:44:44Z"}, {"author": "Anders K\u00f6lligan", "text": "IONOS is also a cloud provider.
", "time": "2025-07-25T07:45:33Z"}, {"author": "Jason Livingood", "text": "great comment from Lorenzo! Excellent point on the 5% loss.
", "time": "2025-07-25T07:46:32Z"}, {"author": "Mike Blanche", "text": "@Anders - but traffic going from HE (a global backbone provider) to IONOS and then back to HE makes absolutely no sense
", "time": "2025-07-25T07:48:00Z"}, {"author": "Mike Blanche", "text": "which made me wonder what that traceroute graph was showing
", "time": "2025-07-25T07:48:18Z"}, {"author": "Florian Obser", "text": "the k.root-servers.org traceroute didn't make sense either. AS25152 is k.root-servers.net, k.root-servers.org is operated by AS3333. also, if you trace to AS25152 you are not going to see 3 hops inside of that AS.
", "time": "2025-07-25T07:50:53Z"}, {"author": "Mike Blanche", "text": "yeah, typo maybe. k-root is anycast from AS25152. If HE is starlink's preferred transit provider I'd expect almost every traceroute to go directly 6939->25152
", "time": "2025-07-25T07:54:44Z"}, {"author": "Jason Livingood", "text": "I pinged a buddy at Starlink to ask if he could come to IETF-124 & present in some WG.
", "time": "2025-07-25T07:57:39Z"}, {"author": "Anders K\u00f6lligan", "text": "@Mike the plot also says \"Most Frequent ASN\". Now I am unsure if this means \"most frequent next hop\" or \"most frequent AS for that number of hops\" probably the latter?
", "time": "2025-07-25T08:08:53Z"}, {"author": "Lucas Pardue", "text": "Thanks for this work and for reaching out to implementers
", "time": "2025-07-25T08:28:26Z"}, {"author": "Alan Frindell", "text": "no proxygen/mvfst :(
", "time": "2025-07-25T08:32:17Z"}, {"author": "Mirja K\u00fchlewind", "text": "mvfst has so many dependencies and is much harder to use... at least in my out-dated experience.
", "time": "2025-07-25T08:33:18Z"}, {"author": "Lucas Pardue", "text": "Its splitting hairs but you can't test \"QUIC-only\" because QUIC delegates responsibility for stream or datagram data to applications.
\nYou can have a \"minimal application layer\" like HTTP/0.9 as shown on slide 6. But the \"QUIC-only\" name is a minsomer - its just a different minimal application layer.
", "time": "2025-07-25T08:33:47Z"}, {"author": "Alan Frindell", "text": "For H3 measurements, just pull and build proxygen, which will pull in mvfst and the deps.
\nWe do serve a lot of traffic, despite the github stars :)
", "time": "2025-07-25T08:34:33Z"}, {"author": "Nick Banks", "text": "No msquic either :(
", "time": "2025-07-25T08:35:22Z"}, {"author": "Alan Frindell", "text": "Nick: :(
", "time": "2025-07-25T08:35:30Z"}, {"author": "Mirja K\u00fchlewind", "text": "okay, yes, we actually wanted to modify the stack but if you just want to test it, it should be easy enough
", "time": "2025-07-25T08:35:54Z"}, {"author": "Mirja K\u00fchlewind", "text": "come to the mic and tell them to come back next time with more data
", "time": "2025-07-25T08:36:19Z"}, {"author": "Alan Frindell", "text": "There's also an http performance testing tool in the proxygen repo: httperf2: https://github.com/facebook/proxygen/tree/main/proxygen/httpclient/samples/httperf2
", "time": "2025-07-25T08:36:53Z"}, {"author": "Nick Banks", "text": "@lucas I agree you always need a protocol on top, but my quic perf draft does outline a minimal one that in practice tests only quic
", "time": "2025-07-25T08:37:09Z"}, {"author": "Lucas Pardue", "text": "need more precise definitions of what \"only QUIC\" means. Is it eg.\"minimal application space processing\" or is it \"minimal application protocol framing on top of QUIC streams\" or ...
", "time": "2025-07-25T08:38:55Z"}, {"author": "Nick Banks", "text": "I agree. That's why a spec is nice
", "time": "2025-07-25T08:39:14Z"}, {"author": "Alan Frindell", "text": "how many parallel requests in the test setup?
", "time": "2025-07-25T08:39:23Z"}, {"author": "Lucas Pardue", "text": "agree Nick
", "time": "2025-07-25T08:39:43Z"}, {"author": "Nick Banks", "text": "I was assuming a single request...
", "time": "2025-07-25T08:39:53Z"}, {"author": "Alan Frindell", "text": "The CPU graph was the sum of all cores, but not sure how you would use more than one if only one request at a time
", "time": "2025-07-25T08:40:17Z"}, {"author": "Lucas Pardue", "text": "FWIW if this was a performance test using quiche open source tools, take the results with a pinch of salt because those are not optimized for perf
", "time": "2025-07-25T08:40:30Z"}, {"author": "Nick Banks", "text": "Msquic supports QUIC (crypto) on one core and UDP on another
", "time": "2025-07-25T08:41:00Z"}, {"author": "Alan Frindell", "text": "yeah but they have 8 cores in the test host
", "time": "2025-07-25T08:41:19Z"}, {"author": "Nick Banks", "text": "It can effectively double throughout for these types of tests
", "time": "2025-07-25T08:41:20Z"}, {"author": "Nick Banks", "text": "Throughput*
", "time": "2025-07-25T08:41:32Z"}, {"author": "Alan Frindell", "text": "we don't serve a lot of 8gb http responses, though more now that folks like to download models
", "time": "2025-07-25T08:41:44Z"}, {"author": "Nick Banks", "text": "Yeah, it's not a default setting. Only stuff like SMB would use it
", "time": "2025-07-25T08:42:58Z"}, {"author": "Lucas Pardue", "text": "thanks for the perf talk though!
", "time": "2025-07-25T08:44:16Z"}, {"author": "Chris Box", "text": "Does this mean pacing individual packets, or sending packets in small bursts with pacing between bursts? The former is terrible for Wi-Fi.
", "time": "2025-07-25T08:45:27Z"}, {"author": "Chris Box", "text": "Ok so there are small bursts
", "time": "2025-07-25T08:49:48Z"}, {"author": "Lucas Pardue", "text": "the authors of the pacing paper might be interested to know that quiche logs a custom \"send_at_time\" filed in the qlog packet sent event, that could be used to link the packet number between internal view and on wire reality
\n", "time": "2025-07-25T08:50:10Z"}, {"author": "Dave Plonka", "text": "@Chris - good point though. It's interesting to see how this would work where a network segment was on shared radio media.
", "time": "2025-07-25T08:50:35Z"}, {"author": "Dave Plonka", "text": "thanks @lucas
", "time": "2025-07-25T08:51:14Z"}, {"author": "Gorry Fairhurst", "text": "I wonder how much pacing is benefcial, it's nice to see all the pacers do quite well at avoiding bursts ~ IW, and short bursts may actually be better for multiplexing?
", "time": "2025-07-25T08:59:13Z"}]