The distinction between CQ completion and remote delivery makes the perftest comparison much clearer. One reproducibility detail that could help interpret the 7.7 µs event-vs-poll median delta is whether the submitting thread and IRQ/NIC processing were CPU-pinned; even on Soft-RoCE, scheduler placement can shift that wake-up cost. Your caveat about the three-vCPU VM and noisy maxima already points to why it matters.