Hello!
I have a somewhat unusual question related to the new “Packet Trimming” feature of the Broadcom Tomahawk5 chipset. In a nutshell, the idea is quite simple: during congestion on a switch, instead of dropping the entire packet, only the payload is dropped, and the headers are forwarded to the destination (this is one of the main ideas of the UltraEthernet Specification).
My question is: how does the Mellanox ConnectX-7 process such packets? I am considering running some tests with this scenario, but first it would be great to understand how the NIC processes packet headers without a payload in a RoCEv2 RDMA scenario.
I would appreciate any answers or relevant links.
Thanks in advance,
Roman
Hi Roman,
Packet trimming is supported on NVIDIA end-to-end solution based on Spectrum switches with our Spectrum-X solution.
In Spectrum-X E2E solution our NIC HW has the support of ingesting these packets and react on them per the functionality defined in the E2E solution.
Regards,
Yaniv
Hi Yaniv,
Thank you for your answer! Could you please share the link to the documentation for the E2E solution you mentioned? Also, just to confirm, does it work with the ConnectX-7 NIC?
Regards,
Roman
Thank you for the link. I have reviewed the white papers on the NVIDIA Spectrum-X Platform, but I have not found any mention of trimming feature. Additionally, all the materials seem to refer to the BlueField NIC, whereas my question specifically concerns the ConnectX-7. Could you please direct me to documentation that describes the trimming feature in the context of NVIDIA? Does this feature work with the ConnectX-7 NIC?
Hi,
It is supported on Spectrum-X solution. As such it is enabled (under the hood) by the BlueField device.
Regards,
Yaniv