[PATCH v8 00/14] Multi-Gen LRU Framework

[Linus Torvalds]

On Tue, Mar 8, 2022 at 4:15 PM Yu Zhao <yuzhao at google.com> wrote:
>
> This sounds self-serving: our data centers want them, so I had to try.

Heh. I'm not opposed to putting them back in, but if/when we merge the
multi-gen LRU code, I really want people to be all testing the same
thing.

I also think that if we put them back in, that should come with

 (a) performance numbers for the different cases

 (b) hard guidance of what the numbers should be, and under what
circumstances (ie giving the user enough information that he *can*
answer the question for his configuration)

 (c) some thought about perhaps making them possibly more dynamic than
a hardcoded build-time value (assuming the numbers show that it's
worth doing in the first place, of course)

so I think that the support for the concept can/should be left in, but
I think that kind of fancy "I want more generations or fewer
tiers-per-generation because of XYZ" needs to be a separate issue with
more explanation from the initial "This multi-gen LRU gives better
performance" merge.

Because as-is, I don't think those config options had nearly enough
information associated with them to merit them existing.

                  Linus