Connectx-6 issue with RoCE v2

We have a rping application running on a Xilinx device to send RoCE v2 package to a Connectx-6 card, the Xilinx device is the server and Connectx-6 card is the client. The cnp request is processed no problem, the RC connection is created, but after that the all the RDMA request/response to the Connectx-6 card are not recognized, the received packet counter show the correct number, but the packet disappeared after that, check the hw_counters, find local_ack_timeout_err, slow_restart_cnps, and rp_cnp_handled also have a larger value.
To compare, we also do the self loop back test on the Mellanox card and it works no problem.
Compared the captured pcapng file, we noticed the ECN field in IP head is different, Xilinx device always set it as 01 (ECT(1)), and Connectx-6 card set it 10 (ECT(0)). Is this the reason that the received packet got ignored? If it is, is there a way to solve it?

NVIDIA has it’s own ROCE CC algorithm DCQCN on CX6. And ROCE is EtoE solution, need work on NVIDIA End to End device.

If you want developing CC algorithm you can contact NVIDIA pre-sales support.

Hi Xiaofengl,
Thank you so much for your quick response! Do you mean Connectx-6 cannot works other venders RoCE device? Could you please give more information?
Since the original demo is using a Connectx-4 card, do you think we have a chance to avoid this issue with a Connectx-4 card? We just want to validate the 100G data transmitting from Xilinx device.

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.