Hi Rik,

On Thu, Oct 20, 2022 at 11:28:16AM -0400, Rik van Riel wrote:
> On Thu, 2022-10-20 at 13:07 +0800, Huang, Ying wrote:
> > 
> > Nathan Chancellor <nathan@kernel.org> writes:
> > > 
> > > For what it's worth, I just bisected a massive and visible
> > > performance
> > > regression on my Threadripper 3990X workstation to commit
> > > f35b5d7d676e
> > > ("mm: align larger anonymous mappings on THP boundaries"), which
> > > seems
> > > directly related to this report/analysis. I initially noticed this
> > > because my full set of kernel builds against mainline went from 2
> > > hours
> > > and 20 minutes or so to over 3 hours. Zeroing in on x86_64
> > > allmodconfig,
> > > which I used for the bisect:
> > > 
> > > @ 7b5a0b664ebe ("mm/page_ext: remove unused variable in
> > > offline_page_ext"):
> > > 
> > > Benchmark 1: make -skj128 LLVM=1 allmodconfig all
> > >   Time (mean ± σ):     318.172 s ±  0.730 s    [User: 31750.902 s,
> > > System: 4564.246 s]
> > >   Range (min … max):   317.332 s … 318.662 s    3 runs
> > > 
> > > @ f35b5d7d676e ("mm: align larger anonymous mappings on THP
> > > boundaries"):
> > > 
> > > Benchmark 1: make -skj128 LLVM=1 allmodconfig all
> > >   Time (mean ± σ):     406.688 s ±  0.676 s    [User: 31819.526 s,
> System: 16327.022 s]
> > >   Range (min … max):   405.954 s … 407.284 s    3 run
> > 
> > Have you tried to build with gcc?  Want to check whether is this
> > clang
> > specific issue or not.
> 
> This may indeed be something LLVM specific. In previous tests,
> GCC has generally seen a benefit from increased THP usage.
> Many other applications also benefit from getting more THPs.

Indeed, GCC builds actually appear to be slightly faster on my system now,
apologies for not trying that before reporting :/

7b5a0b664ebe:

Benchmark 1: make -skj128 allmodconfig all
  Time (mean ± σ):     355.294 s ±  0.931 s    [User: 33620.469 s, System: 6390.064 s]
  Range (min … max):   354.571 s … 356.344 s    3 runs

f35b5d7d676e:

Benchmark 1: make -skj128 allmodconfig all
  Time (mean ± σ):     347.400 s ±  2.029 s    [User: 34389.724 s, System: 4603.175 s]
  Range (min … max):   345.815 s … 349.686 s    3 runs

> LLVM showing 10% system time before this change, and a whopping
> 30% system time after that change, suggests that LLVM is behaving
> quite differently from GCC in some ways.

The above tests were done with GCC 12.2.0 from Arch Linux. The previous LLVM
tests were done with a self-compiled version of LLVM from the main branch
(16.0.0), optimized with BOLT [1]. To eliminate that as a source of issues, I
used my distribution's version of clang (14.0.6) and saw similar results as
before:

7b5a0b664ebe:

Benchmark 1: make -skj128 LLVM=/usr/bin/ allmodconfig all
  Time (mean ± σ):     462.517 s ±  1.214 s    [User: 48544.240 s, System: 5586.212 s]
  Range (min … max):   461.115 s … 463.245 s    3 runs

f35b5d7d676e:

Benchmark 1: make -skj128 LLVM=/usr/bin/ allmodconfig all
  Time (mean ± σ):     547.927 s ±  0.862 s    [User: 47913.709 s, System: 17682.514 s]
  Range (min … max):   547.429 s … 548.922 s    3 runs

> If we can figure out what these differences are, maybe we can
> just fine tune the code to avoid this issue.
> 
> I'll try to play around with LLVM compilation a little bit next
> week, to see if I can figure out what might be going on. I wonder
> if LLVM is doing lots of mremap calls or something...

If there is any further information I can provide or patches I can test,
I am more than happy to do so.

[1]: https://github.com/llvm/llvm-project/tree/96552e73900176d65ee6650facae8d669d6f9498/bolt

Cheers,
Nathan