While TRex is able to serve 64B frames on 40GE NICs quite well, its performance on 100GE NICs is not consistent. Sometimes it is able to reach line rate, other times some frames remain unsent, in numbers that affect even PDR (MRR is fine). Trending graph [0] shows (as expected) clx is affected more than icx.
Only some highest performing VPP configurations (at 4 cores) are affected, visible as large standard deviation of iterative results [1].

Not sure how to fix this. Perhaps some work on TRex calibration (tweaking number of workers or perhaps size of queues or even time delays) may help.

[0] https://csit.fd.io/trending/#eNrNlMFOwzAMhp-mXJClNrSUCwdG32NKHW-tSDuTZNPG05NWA29aN4mBgEMSS7_j3_4UxYeVo7kn-5gUs6ScJapsTdySu6fbeARHW1A9tLiFLE2XpDijhyzFVzBsXqDlHO7zGoBCM8Rx1doThCWwcUMd9TzUMetwXFQkbnYinbeSC9qRlhvRUaRA_sDnpCdJXDjdkW_fSLLjGKJjpCLSsUHY8YG2H7OsxowvA7Tqk1-Ma-NRW8q6TqN1_S9itOoCxcnO_gDmhnlgifZkNL1ZfKDMcQ-TdcDmAr9Y7Ty-aY9r6Y2tXAssx-8jm3gNP45s2uNfIyuqm37luvHvK6p34T1a2w
[1] https://csit.fd.io/report/#eNrtlM-OgjAQxp-GvWzGlAFkL3vQ5T1MLeNiUqA7rUR9eqshOxBjPHnz0knz-zr_vqQ-9EwbT_Y7KdZJuU6w3NfxSLLVZwxsPWZqCYNzgNlCITBZ0p4AOzD2CKlSv4Qupa9UmT_Qww4swjLfQm6AQhNvTgfTQMt8TYs_Y9q7GkLrQxAaK8_IQCxw1pLIXHOaaB41KnrNpOWBRSGB_KSXyTwi2bFuye_PJLo4vnATFywoN_Pk4eQmdFxSWd0UL3Sjq99uPHVjXNLr3XBvN5674f7dKKqPruf29mMV1QWVwkKL