On Monday, March 20, 2017 3:40:46 AM GMT Lanlan Pan wrote: > At NDSS there is a question that "why not directly use AS number" ? client > subnet can be maped into AS number, which is used for bgp route at network > topology. > > My answer was that AS4134 cover multiple provinces in china, from south > China to north China, 2000+ kilometers physical distance, also, with long > latency at network topology. > > As the maxmind ip geo database showed in the slide, my key point is that, > the "country, province, isp" information can offer the same "network > topology" information like client subnet.
three eyeball-sets in the city of san francisco might be reachable only through either los angeles, or else sacramento, or else salt lake city. congestion effects on the long haul links will loom larger than speed-of- photons-in-fibre in the performance differences a CDN might measure if they served the same content to the same user from a data center in los angeles vs. sacramento vs. salt lake city. > We tell Authorative servers that, "I want to know what is most satisfied ip > address for clients from CHINA, BEIJING, TELECOM at network topology". > > But not "I want to know what is the nearest ip address for clients from > CHINA, BEIJING, TELECOM at physical topology". in my example all three eyeball-sets have the same last-mile autnum. nothing is reliably predictive of last mile performance, except prior experience serving similar content to the same client-subnet. estimating the client's subnet-size as /24 does some harm to accuracy. estimating it at /28 does some harm to state-load and state-churn. but anyway you have to serve it before you know-- the routing table, the AS path, and the geo-loc are each nonpredictive, though each for different reasons. i am not an ECS fan. far from it! but for what it purports to do, only the client's actual subnet will "work" in the live-fire exercise i call "home". vixie _______________________________________________ DNSOP mailing list DNSOP@ietf.org https://www.ietf.org/mailman/listinfo/dnsop
