Very interesting link, thanks. Obviously your use case goes way beyond running a standard Docker installation. It sounds like you really want the kernel to schedule containers instead of processes -- which it doesn't really do by default, hence my comment. Perhaps it shouldn't be surprising, that you're able to get these clear performance gains from a highly optimized special-purpose scheduler. Still, I was a bit surprised. :)

However, that's at the OS level. What can realistically be done at the hardware level? It must be possible in theory to design a CPU that's better at this kind of context switching, but I don't know if a new "computer company" really wants to go there.