Re: [PATCH 1/1] mm/khugepaged: reduce process visible downtime by pre-zeroing hugepage

[Lance Yang]

Another thought suggested by Bang Li is that we record which pte is none
in hpage_collapse_scan_pmd. Then, before acquiring the mmap_lock (write mode),
we will pre-zero pages as needed.

What do you think?

Thanks,
Lance

On Tue, Mar 12, 2024 at 9:55 PM Lance Yang <ioworker0@xxxxxxxxx> wrote:
>
> On Tue, Mar 12, 2024 at 9:19 PM David Hildenbrand <david@xxxxxxxxxx> wrote:
> >
> > On 12.03.24 14:09, Lance Yang wrote:
> > > Hey David,
> > >
> > > Thanks for taking time to review!
> > >
> > > On Tue, Mar 12, 2024 at 12:19 AM David Hildenbrand <david@xxxxxxxxxx> wrote:
> > >>
> > >> On 08.03.24 08:49, Lance Yang wrote:
> > >>> The patch reduces the process visible downtime during hugepage
> > >>> collapse. This is achieved by pre-zeroing the hugepage before
> > >>> acquiring mmap_lock(write mode) if nr_pte_none >= 256, without
> > >>> affecting the efficiency of khugepaged.
> > >>>
> > >>> On an Intel Core i5 CPU, the process visible downtime during
> > >>> hugepage collapse is as follows:
> > >>>
> > >>> | nr_ptes_none  | w/o __GFP_ZERO | w/ __GFP_ZERO  |  Change |
> > >>> --------------------------------------------------—----------
> > >>> |      511      |     233us      |      95us      |  -59.21%|
> > >>> |      384      |     376us      |     219us      |  -41.20%|
> > >>> |      256      |     421us      |     323us      |  -23.28%|
> > >>> |      128      |     523us      |     507us      |   -3.06%|
> > >>>
> > >>> Of course, alloc_charge_hpage() will take longer to run with
> > >>> the __GFP_ZERO flag.
> > >>>
> > >>> |       Func           | w/o __GFP_ZERO | w/ __GFP_ZERO |
> > >>> |----------------------|----------------|---------------|
> > >>> | alloc_charge_hpage   |      198us     |      295us    |
> > >>>
> > >>> But it's not a big deal because it doesn't impact the total
> > >>> time spent by khugepaged in collapsing a hugepage. In fact,
> > >>> it would decrease.
> > >>
> > >> It does look sane to me and not overly complicated.
> > >>
> > >> But, it's an optimization really only when we have quite a bunch of
> > >> pte_none(), possibly repeatedly so that it really makes a difference.
> > >>
> > >> Usually, when we repeatedly collapse that many pte_none() we're just
> > >> wasting a lot of memory and should re-evaluate life choices :)
> > >
> > > Agreed! It seems that the default value of max_pte_none may be set too
> > > high, which could result in the memory wastage issue we're discussing.
> >
> > IIRC, some companies disable it completely (set to 0) because of that.
> >
> > >
> > >>
> > >> So my question is: do we really care about it that much that we care to
> > >> optimize?
> > >
> > > IMO, although it may not be our main concern, reducing the impact of
> > > khugepaged on the process remains crucial. I think that users also prefer
> > > minimal interference from khugepaged.
> >
> > The problem I am having with this is that for the *common* case where we
> > have a small number of pte_none(), we cannot really optimize because we
> > have to perform the copy.
> >
> > So this feels like we're rather optimizing a corner case, and I am not
> > so sure if that is really worth it.
> >
> > Other thoughts?
>
> Another thought is to introduce khugepaged/alloc_zeroed_hpage for THP
> sysfs settings. This would enable users to decide whether to avoid unnecessary
> copies when nr_ptes_none > 0.
>
> >
> > --
> > Cheers,
> >
> > David / dhildenb
> >