irq cpu binding

[Slawa Olhovchenkov]

On Sun, Mar 29, 2015 at 10:01:06AM -0700, Adrian Chadd wrote:

> Well, have you seen how fast the windows network apps can get? Not the
> gui torrent clients, but the really serious HTTP intermediary stuff
> and message passing stuff?
> 
> The model works and scales quite well.. :0

"how fast"... What is "fast"? How we calculate speed?
In what enviroment (dual or single socket)?
nginx is acceptable fast. FreeBSD is acceptable fast.
But I am need less overhead [for multi-socket system].
I am may be wrong.
I am use dual-socket systems.
And when I am think about pinning to CPU I am think about elimination
transfer context by QPI link.
May be for single-socket systems this is will be less impact.
May be double of interrupt/network/TCP processing thread gain lock
congestion and eliminate all win.
You have other opinion?

About NUMA. I have expirense with Tilera. Perfomance with NUMA off
(interleave memory) will be better.