Re: [PATCH net-next v2 0/9] xsk: improvement performance in copy mode

On Tue, Aug 26, 2025 at 8:29 AM Jakub Kicinski [off-list ref] wrote:

On Tue, 26 Aug 2025 08:01:03 +0800 Jason Xing wrote:

quoted

On Tue, Aug 26, 2025 at 1:44 AM Jakub Kicinski [off-list ref] wrote:

quoted

On Mon, 25 Aug 2025 21:53:33 +0800 Jason Xing wrote:

quoted

copy mode:   1,109,754 pps
batch mode:  2,393,498 pps (+115.6%)
xmit.more:   3,024,110 pps (+172.5%)
zc mode:    14,879,414 pps

I've asked you multiple times to add comparison with the performance
of AF_PACKET. What's the disconnect?

Sorry for missing the question. I'm not very familiar with how to run the
test based on AF_PACKET. Could you point it out for me? Thanks.

I remember the very initial version of AF_XDP was pure AF_PACKET. So
may I ask why we expect to see the comparison between them?

Pretty sure I told you this at least twice but the point of AF_XDP
is the ZC mode. Without a comparison to AF_PACKET which has similar
functionality optimizing AF_XDP copy mode seems unjustified.

Oh, I see. Let me confirm again that you expect to see a demo like the
copy mode of AF_PACKET v4 [1] and see the differences in performance,
right?

If AF_PACKET eventually outperforms AF_XDP, do we need to reinvent the
copy mode based on AF_PACKET?

And if a quick/simple implementation is based on AF_PACKET, it
shouldn't be that easy to use the same benchmark to see which one is
better. That means inventing a new unified benchmark tool is
necessary?

[1]: https://lore.kernel.org/all/20171031124145.9667-1-bjorn.topel@gmail.com/ (local)

Thanks,
Jason