The progression is visible for both AVF and VFIO_PCI variants of tests, on all examined platforms (64b, 2t1c, xxv710):
+ 2n-skx:
+ avf-eth-l2bdscale10kmaclrn: +19.5 +-1.9
+ eth-l2bdscale10kmaclrn: +13.1 +-1.9
+ 2n-clx:
+ avf-eth-l2bdscale10kmaclrn: +18.1 +-2.5
+ eth-l2bdscale10kmaclrn: +13.8 +-2.5
+ 3n-skx:
+ avf-eth-l2bdscale10kmaclrn: +16.6 +-1.4

Preliminary investigation found bihash optimization [0] to be the primary cause, but performance difference suggests it is not the only contributing cause.

The 100k tests are affected to almost the same degree. Strangely enough, 1m test are either not affected, or suffering from an unrelated regression making the overall performance difference negligible.

[0] https://gerrit.fd.io/r/c/vpp/+/26661