Stephen Hemminger wrote:

These loopback benchmarks are often more sensitive to scheduler than networking
changes.

Just ran a test with real NICs which show the same issues. I guess I need to
get familiar with the network stack and start hacking on it. Sigh.