Example trending graph: [0].
Affects all scales, all tested testbeds, both MRR and PDR.
Affects UDP more than TCP. Typical regression magnitudes are 13% for UDP and 6% for TCP.

Bisect found [1] to be the cause, which makes sense as it explicitly adds more computation to slow path (in order to support multiple VRFs).
The question is whether VPP code can be better optimized.

[0] http://csit.fd.io/trending/#eNqNkMFqwzAMhp_GuRSBrTjxqYe1eY_hxuoStqbCdkO6p68DATWHwS4y4vv1_5JTvkf6TPRzVM1JuZNCN4ZSVP1xKM_MDDhB-l4Amy9CNssyO6PBz1cY2UJrL2B6oDyU7hEYJp-tpQCD0WiB2xpSaxuD0HOCW4yrN55X7_DIuyAhPDyF_BUvch_Ji77sIShTegv535YyfY3-Rmn8JbEo9wrvy98JMv0-Nz_5jW6nu041XZWyj_mIGhE0AuJB6yLbSkVT2JgDU-_YCx58coQ
[1] https://gerrit.fd.io/r/c/vpp/+/35429