Re: [RFC] mm/swap.c: Enable promotion of unmapped MGLRU page cache pages

[Donet Tom]

On 1/15/25 21:39, Gregory Price wrote:
> On Wed, Jan 15, 2025 at 06:06:25AM -0600, Donet Tom wrote:
> ... snip ...
>
> Thank you for taking the time to do this, I don't have enough background
> with MGLRU to have done this quickly.  I'll pull this onto my branch and
> carry it if you don't mind so we can keep things tracked.
>
> I'll send you an updated RFC before i send out v4 and add appropriate tags.
Sure. Thank you.
> > This difference also impacts read latency:
> >
> > For MGLRU, the first read shows higher latency due to the combined
> > overhead of accessing a lower tier and performing promotion.
> >
> > For LRU, the first 3–4 reads typically exhibit lower latency since
> > promotion does not occur immediately.
> Do you have a thought on a good test we can use to compare these
> strategies?
I am also thinking on the same. I will come back with some tests which
can compare these strategies.
> We decided against promotion on first-access because there are many
> easy-to-imagine scenarios where that will clearly harm performance.
Yes, I agree. After the first access, it gets promoted, but if those pages
are not accessed afterward, it will become an overhead.
I will check if we have any mechanism to restrict it.
> We're planning to do some workload testing soon so we can get actual
> benefit numbers.
Could you please let me know what workload you are planning?
I can try it on my system as well.
> > +promo_candid:
> > +	if (!folio_test_isolated(folio) &&
> > +		(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) &&
> > +		numa_pagecache_promotion_enabled) {
> I am considering putting this in some inline wrapper with some likely()
> tags to clean this up a bit and optimize the fall-through cases since
> i've seen some measurable differences when left as-is.
Sounds good to me.
> Thoughts on this are welcome
>
> > +		memcg = folio_memcg(folio);
> > +		if (memcg) {
> Also curious, why only promote when the folio is a member of a memcg?

I missed non memcg pages. I will send a V2 with that.

~Donet

> ~Gregory