Conducting traffic congestion testing using Broadcom's TD3 environment, from a 100G port to a 25G port.
Client: ib_write_bw 11.4.1.1 -d mlx5_0 -s 50M -n 100 -p 12123 --report_gbits
Server: ib_write_bw -d mlx5_1 -p 12123 -s 50M -n 100 --report_gbits
Set ECN watermark thresholds to 400/1625/20 (units: cells/cells/percentage).
PFC and CNP frames were observable.
Using identical parameters in NS3 simulation, PFC pause frames were not observed.
Suspicion: The NIC node introduces delay (approximately 55,000 ns) when processing ECN-marked packets to generate CNP. Adding this delay triggered PFC pause frames. However, after modifying the ECN watermarks, the results still do not match the lab data.