Depending on where in the search the trial with unsent packet happens, the test may not fail, just report low critical load, which is worse than failing. For high load, unsent packets are expected as a symptom of duration stretching. But at not so high load, unsent packets count as loss that has large consequences due to fitting functions. Here [0] the fifth trial had 326 unsent packets. I do not think the AVF driver on VPP side is guilty, this is probably a trex issue. Infra workaround by using more reliable NIC is preferred, but if that is not viable, soak tests could include additional workarounds to filter out bad trials.

[0] https://s3-logs.fd.io/vex-yul-rot-jenkins-1/csit-vpp-perf-report-iterative-2410-2n-icx/60/log.html.gz#s1-s1-s1-s4-s2-t1-k2-k9-k13