Most of test on 3n-skx show ~-5% difference of CSIT code, but with high stdev making it a noise. Still, with pure noise one would expect occasional positive difference, but that does not happen.

One specific test case shows quite small stdev for some reason:
imix-2t1c-avf-ethip4ipsec4tnlsw-ip4base-int-aes256gcm

With small stdev, the performance difference strongly suggests there was a CSIT code change affecting the performance.

The small stdev could be just luck, but this test is expected to have the best performance out of all ipsec tests run for report, so perhaps the smaller stdev is caused by that (as higher offered load means VPP is less sensitive to time distribution of packets).

Therefore it is possible that manual bisect in CSIT code (with a workaround to keep the number of TRex workers constant) will identify a real cause in CSIT.