> As soon as you get the hardware to handle TCP reassembly and just wake the kernel up once per few megabytes of data sent/received, things scale well again.

Not a generic solution, because it increases latency.

Which you could avoid by doing everything in user or kernel space.