Re: [RFC] Using "page credits" as a solution for common thrashing scenarios

[Eyal Lotem <eyal.lotem@gmail.com>]

> The real problematic details are in the data structures.
>
> How to organize the pages in-kernel?
>
> How does the kernel find the pages with the lowest
> score?  In the memory zone where memory needs
> to be freed...
>
> How can the kernel efficiently change the score of
> tens of thousands (or even millions) of pages at once,
> moving them to the right LRU list cheaply?
>
> Making page replacement scale to systems with
> hundreds of gigabytes of memory is a necessity, as
> is evicting the right kind of page (filesystem cache
> vs process page, etc).
>
> Within those constraints, it becomes very hard to
> implement a solution along your ideas.

I agree an implementation that reasonably addresses these concerns is far
from trivial.  I wouldn't give up on finding a way to implement it (or a
good approximation of it) so soon.

Before diving into the exact implementation details, I was wondering if I'm
getting some basic fact wrong -- would this solution (if possible to
implement with reasonable performance) truly mitigate this kind of
thrashing?

Eyal