Re: [PATCH v2 2/2] mm/pageblock: remove false sharing in pageblock_flags

From: Alexander Duyck
Date: Sun Aug 30 2020 - 16:32:55 EST


On Sun, Aug 30, 2020 at 3:00 AM Alex Shi <alex.shi@xxxxxxxxxxxxxxxxx> wrote:
>
>
>
> 在 2020/8/20 上午12:50, Alexander Duyck 写道:
> > On Wed, Aug 19, 2020 at 1:11 AM Alex Shi <alex.shi@xxxxxxxxxxxxxxxxx> wrote:
> >>
> >>
> >>
> >> 在 2020/8/19 下午3:57, Anshuman Khandual 写道:
> >>>
> >>>
> >>> On 08/19/2020 11:17 AM, Alex Shi wrote:
> >>>> Current pageblock_flags is only 4 bits, so it has to share a char size
> >>>> in cmpxchg when get set, the false sharing cause perf drop.
> >>>>
> >>>> If we incrase the bits up to 8, false sharing would gone in cmpxchg. and
> >>>> the only cost is half char per pageblock, which is half char per 128MB
> >>>> on x86, 4 chars in 1 GB.
> >>>
> >>> Agreed that increase in memory utilization is negligible here but does
> >>> this really improve performance ?
> >>>
> >>
> >> It's no doubt in theory. and it would had a bad impact according to
> >> commit e380bebe4771548 mm, compaction: keep migration source private to a single
> >>
> >> but I do have some problem in running thpscale/mmtest. I'd like to see if anyone
> >> could give a try.
> >>
> >> BTW, I naturally hate the false sharing even it's in theory. Anyone who doesn't? :)
> >
> > You keep bringing up false sharing but you don't fix the false sharing
> > by doing this. You are still allowing the flags for multiple
> > pageblocks per cacheline so you still have false sharing even after
> > this.
>
> yes, the cacheline false sharing is still there. But as you pointed, cmpxchg level
> false sharing could be addressed much by the patchset.
>
>
> >
> > What I believe you are attempting to address is the fact that multiple
> > pageblocks share a single long value and that long is being used with
> > a cmpxchg so you end up with multiple threads potentially all banging
> > on the same value and watching it change. However the field currently
> > consists of only 4 bits, 3 of them for migratetype and 1 for the skip
> > bit. In the case of the 3 bit portion a cmpxchg makes sense and is
> > usually protected by the zone lock so you would only have one thread
> > accessing it in most cases with the possible exception of a section
> > that spans multiple zones.
> >
> > For the case such as the skip bit and MIGRATE_UNMOVABLE (0x0) where we
> > would be clearing or setting the entire mask maybe it would make more
> > sense to simply use an atomic_or or atomic_and depending on if you are
> > setting or clearing the flag? It would allow you to avoid the spinning
> > or having to read the word before performing the operation since you
> > would just be directly applying an AND or OR via a mask value.
>
> Right that the different level to fix this problem, but narrow the cmpxchg
> comparsion is still needed and helpful.

What I was getting at though is that I am not sure that is the case.
Normally I believe we are always holding the zone lock when updating
the migrate type. The skip flag is a one-off operation that could
easily be addressed by changing the logic to use atomic_and or
atomic_or for the cases where we are updating single bit flags and
setting the mask value to all 1's or all 0's. So adding this extra
complexity which only really applies to the skip bit may not provide
much value, especially as there are a number of possible paths that
don't use the skip bit anyway.