Hundreds of milliseconds per TCP (or UDP) session. The CGNAT boxes are handling hundreds of thousands of flows. And they're part of failover clusters so each flow has to be shipped to the other cluster members and acknowledged before starting the flow. And these things are often overloaded and adding tens of milliseconds delay to each packet passing through. Because whoever heard of an ISP spending more than "just enough" on hardware?
You see, my point is that these milliseconds are negligible in the case of a website like Facebook that takes 20 second to fully load on a reasonably modern machine.
You may already be behind an overloaded CGNAT then. Over IPv6 on Comcast cable, with the Edge browser, Facebook is fully loaded for me in 3 seconds. Surface tablet in power-save mode so this isn't some overpowered desktop monster.
That's very interesting. I've just measured it in private mode, all plugins off, from the moment of clicking login till the time the page load indicator stopped running. 13 full seconds.
I mean, we're probably not talking about a dozen milliseconds out of 20 seconds total, we're talking about extra milliseconds for every request or roundtrip or w/e constituting those 20 seconds. It's an extra cost that scales with the bloat you're decrying, it doesn't become insignificant next to it!
