Comment 5 for bug 1599329

It looks like a bug which may be it's related to the non-netns support for internal ports in ovs. If that's the case, an alternative could be moving to veths, while that is not desirable (specially for router ports).

With a single host, cpu-pegged test run: [1]

I've seen that overhead of using a veth is around +10% in CPU consumption. Please note that the above test tests passing through two veth pairs (so it's 110% * 110% = 121% increase, or ~21% decrease in bandwidth, 20G.5bps to 16Gbps TCP_STREAM or 93k conn/sec TCP_RR to 93k conn/sec TCP_RR)

Talking to some ovs developers, they're willing to fix this (even if it was related to netns usage on internal ports, which is not supported), so I'm continuing kevinbenton's work trying to recreate a bash reproducer.

[1] https://github.com/mangelajo/ovs-experiments/blob/master/veth_vs_intport/veth_vs_intport_speed.sh

[2] https://gist.github.com/kevinbenton/e8a323b61ac6145f9293ef90dbdec0b1