There's also another important detail. Due to delayed ACKs on Windows and battery-preserving TCP stacks on most mobile devices/laptops, some of your packets will be delayed. Even with TCP_NODELAY. Desktop OSX/Linux do not display this behavior.
The only solution I have found, and it's a rather crude one, is to blast the server every 50ms with a 1-byte packet. It is ugly but it works well.
Temporal dithering. Also called FRC - frame rate control.
You might not notice it, but if you are on a TN monitor right now (almost all non-professional monitors) this technique is being used to emulate 24 bit color. Most TN panels can only display 6 bits per channel.
A legitimate usage case would be opening a login popup on another domain that can't be iframe'd. For example OAuth social network login popups. They require window.opener to send back login credentials to the original website.
What prevents the hardware offloading from being used on public interfaces? Even if microsecond-level jitter reduction on public networks is negligible this should reduce CPU load in high PPS deployments, right?
