Stefan Lankes wrote:
> I am not able to reconstruct any performance drawbacks on my system.
> Could you send me your low-level benchmark?

It's attached. As you may see, it's fairly trivial. It just does several
iterations of mbind+touch_all_pages for different power-of-two buffer
sizes. Just replace mbind with madvise in the inner loop if you want to
try with your affinit-on-next-touch.

Which kernels are you using when comparing your next-touch
implementation with Lee's patchset?

Brice