[LSF/MM/BPF TOPIC] Memory Tiering discussion

[LSF/MM/BPF TOPIC] Memory Tiering discussion

[Yang Shi]

Hi folks,

Since last year's LSF/MM we discussed about this topic and I posted my
v3 patchset based on our discussion
(https://lore.kernel.org/linux-mm/1560468577-101178-7-git-send-email-yang.shi@linux.alibaba.com/t/),
but not too much feedback from the community.

Then Dave posted his version
(https://lore.kernel.org/linux-mm/c3d6de4d-f7c3-b505-2e64-8ee5f70b2118@intel.com/).

Both posts share some basic idea and have arbitrary design choices as well.

So I would like to discuss how we should move forward, what pieces we
still miss, what we can combine, etc.

Thanks,
Yang

Re: [LSF/MM/BPF TOPIC] Memory Tiering discussion

[Dave Hansen]

On 1/28/20 11:50 AM, Yang Shi wrote:
> So I would like to discuss how we should move forward, what pieces we
> still miss, what we can combine, etc.

Thanks for proposing this!

In addition to what you've mentioned, I'd really like to close in on
what folks think the first set of patches we _merge_ would look like.
More features than what we've got in the two sets?  Less?

I'd also like to have to rehash a few discussions we've had inside
Intel.  For instance, migrating (demoting) a page cache page is more
expensive than just discarding at reclaim time.  But, it might be
expensive to recreate that page later if we do discard it.  There
doesn't seem to be a single, obviously "right" answer.

Re: [LSF/MM/BPF TOPIC] Memory Tiering discussion

[Yang Shi]

On Tue, Jan 28, 2020 at 4:05 PM Dave Hansen <dave.hansen@intel.com> wrote:
>
> On 1/28/20 11:50 AM, Yang Shi wrote:
> > So I would like to discuss how we should move forward, what pieces we
> > still miss, what we can combine, etc.
>
> Thanks for proposing this!
>
> In addition to what you've mentioned, I'd really like to close in on
> what folks think the first set of patches we _merge_ would look like.
> More features than what we've got in the two sets?  Less?

Good point. We'd better figure out the minimal viable set.

>
> I'd also like to have to rehash a few discussions we've had inside
> Intel.  For instance, migrating (demoting) a page cache page is more

That would be definitely useful.

> expensive than just discarding at reclaim time.  But, it might be
> expensive to recreate that page later if we do discard it.  There
> doesn't seem to be a single, obviously "right" answer.

I agree, so my patchset just deals with anonymous pages. Page cache
seems more tricky.