Per processor network stats have been supported on netvsc and Mellanox NICs for years. They are also supported on Windows.
I routinely use these stats to rule in/out the issues with RSS imbalance due to either NIC not handling RSS correctly, incorrect MSI-X affinity, or not having enough entropy in flows. And yes, I perfectly understand that there are cases that packets received on processor x are processed (in tcp/ip stack) on processor y. But in many cases, there is a correlation between individual processor utilization and the processor where these packets are received on by the NIC. -thanks, ali (some suggested that I do mention on this thread that I am one of the original inventors of RSS. So there you have it. Personally I don't think that telling people "I invented the damn thing and I know what I am talking about" is the right way to handle this.) -----Original Message----- From: Jakub Kicinski <[email protected]> Sent: Thursday, March 14, 2024 12:05 PM To: Haiyang Zhang <[email protected]> Cc: Shradha Gupta <[email protected]>; Shradha Gupta <[email protected]>; [email protected]; [email protected]; [email protected]; [email protected]; Eric Dumazet <[email protected]>; Paolo Abeni <[email protected]>; Ajay Sharma <[email protected]>; Leon Romanovsky <[email protected]>; Thomas Gleixner <[email protected]>; Sebastian Andrzej Siewior <[email protected]>; KY Srinivasan <[email protected]>; Wei Liu <[email protected]>; Dexuan Cui <[email protected]>; Long Li <[email protected]>; Michael Kelley <[email protected]>; Alireza Dabagh <[email protected]>; Paul Rosswurm <[email protected]> Subject: [EXTERNAL] Re: [PATCH] net :mana : Add per-cpu stats for MANA device On Thu, 14 Mar 2024 18:54:31 +0000 Haiyang Zhang wrote: > We understand irqbalance may be a "bad idea", and recommended some > customers to disable it when having problems with it... But it's still > enabled by default, and we cannot let all distro vendors and custom > image makers to disable the irqbalance. So, our host- networking team > is eager to have per-CPU stats for analyzing CPU usage related to > irqbalance or other issues. You need a use case to get the stats upstream. "CPU usage related to irqbalance or other issues" is both too vague, and irqbalance is a user space problem.
