Re: Potential Regression in futex Performance from v6.9 to v6.10-rc1 and v6.11-rc4

[Anders Roxell <anders.roxell@linaro.org>]

On Wed, 4 Sept 2024 at 15:47, David Hildenbrand <david@redhat.com> wrote:
>
> On 04.09.24 12:05, Anders Roxell wrote:
> > On Tue, 3 Sept 2024 at 14:37, David Hildenbrand <david@redhat.com> wrote:
> >>
> >> On 03.09.24 14:21, Anders Roxell wrote:
> >>> Hi,
> >>>
> >>> I've noticed that the futex01-thread-* tests in will-it-scale-sys-threads
> >>> are running about 2% slower on v6.10-rc1 compared to v6.9, and this
> >>> slowdown continues with v6.11-rc4. I am focused on identifying any
> >>> performance regressions greater than 2% that occur in automated
> >>> testing on arm64 HW.
> >>>
> >>> Using git bisect, I traced the issue to commit
> >>> f002882ca369 ("mm: merge folio_is_secretmem() and
> >>> folio_fast_pin_allowed() into gup_fast_folio_allowed()").
> >>
> >> Thanks for analyzing the (slight) regression!
> >>
> >>>
> >>> My tests were performed on m7g.large and m7g.metal instances:
> >>>
> >>> * The slowdown is consistent regardless of the number of threads;
> >>>      futex1-threads-128 performs similarly to futex1-threads-2, indicating
> >>>      there is no scalability issue, just a minor performance overhead.
> >>> * The test doesn’t involve actual futex operations, just dummy wake/wait
> >>>      on a variable that isn’t accessed by other threads, so the results might
> >>>      not be very significant.
> >>>
> >>> Given that this seems to be a minor increase in code path length rather
> >>> than a scalability issue, would this be considered a genuine regression?
> >>
> >> Likely not, I've seen these kinds of regressions (for example in my fork
> >> micro-benchmarks) simply because the compiler slightly changes the code
> >> layout, or suddenly decides to not inline a functions.
> >>
> >> Still it is rather unexpected, so let's find out what's happening.
> >>
> >> My first intuition would have been that the compiler now decides to not
> >> inline gup_fast_folio_allowed() anymore, adding a function call.
> >>
> >> LLVM seems to inline it for me. GCC not.
> >>
> >> Would this return the original behavior for you?
> >
> > David thank you for quick patch for me to try.
> >
> > This patch helped the original regression on v6.10-rc1, but on current mainline
> > v6.11-rc6 the patch does nothing and the performance is as expeced.
>
> Just so I understand this correctly:
>
> It fixed itself after v6.11-rc4, but v6.11-rc4 was fixed with my patch?

I had to double check and no, on v6.11-rc4 with or without your patch
I see the 2% regression.

Cheers,
Anders

>
> If that's the case, then it's really the compiler deciding whether to
> inline or not, and on v6.11-rc6 it decides to inline again.
>
> --
> Cheers,
>
> David / dhildenb
>