ConnectX-5 Hierarchical QOS hardware offload (HTB)

I’ve been testing HTB offload with a ConnectX-5 MCX512A (2x25G) NIC and getting good performance results up to the limit of 256 leaf nodes.

I saw there where previous commits of the mlx5_core driver where the MLX5E_QOS_MAX_LEAF_NODES constant was defined as 1024 so I recompiled a recent kernel with that higher value but still hit the 256 node limit in the card firmware instead of at the driver.

Looking in the release notes for the MLNX_EN driver I see it says under the Packet Pacing section “Up to 512 send queues are supported”.

Is there a definitive number of queues supported by the hardware and is this a firmware limitation and could it be increased?

Are there different limits on the ConnectX-6 or ConnectX-7 cards?

1 Like


Have you found an answer to your question? ConnectX-6 or 7 allows for more than 256 leaf nodes?

I want to test HTB offload with a bluefield2. Can you give me some help.