Re: thunderbolt_net performance and receive-side-scaling (RSS)

Re: thunderbolt_net performance and receive-side-scaling (RSS)

[Mika Westerberg]

Hi,

On Thu, Mar 05, 2026 at 09:56:26AM +0100, Oddbjørn Kvalsund wrote:
> Hi,
> 
> I am seeing fairly good network performance between a 2018 Dell XPS15 laptop
> and a 2021 MacBook using Thunderbolt and the thunderbolt_net driver, with jumbo
> frames enabled on both machines (iperf3 results below). However, while iperf3
> is running, one core on the XPS15 is at 100% utilization due to the ksoftirqd
> process. This makes me wonder whether additional performance could be achieved
> if the load were distributed across multiple cores.
> 
> From what I can tell, this might relate to receive-side scaling (RSS), which
> does not appear to be implemented in the thunderbolt_net driver. Would adding
> RSS support be worthwhile, or are there architectural limitations that would
> make such an optimization ineffective?

I'm not familiar with that but feel free to experiment and if you find that
it provides better throughput please provide patches :)

USB4NET spec can be downloaded from USB-IF, it's part of USB4 spec.