Re: [mm] f35b5d7d67: will-it-scale.per_process_ops -95.5% regression

[Andrew Morton]

On Thu, 01 Dec 2022 15:29:41 -0500 Rik van Riel <riel@xxxxxxxxxxx> wrote:

> On Thu, 2022-12-01 at 19:33 +0100, Thorsten Leemhuis wrote:
> > Hi, this is your Linux kernel regression tracker.
> > 
> > On 28.11.22 07:40, Nathan Chancellor wrote:
> > > Hi Rik,
> > 
> > I wonder what we should do about below performance regression. Is
> > reverting the culprit now and reapplying it later together with a fix
> > a
> > viable option? Or was anything done/is anybody doing something
> > already
> > to address the problem and I just missed it?
> 
> The changeset in question speeds up kernel compiles with
> GCC, as well as the runtime speed of other programs, due
> to being able to use THPs more. However, it slows down kernel
> compiles with clang, due to ... something clang does.
> 
> I have not figured out what that something is yet.
> 
> I don't know if I have the wrong version of clang here,
> but I have not seen any smoking gun at all when tracing
> clang system calls. I see predominantly small mmap and
> unmap calls, and nothing that even triggers 2MB alignment.

2.8% speedup for gcc is nice.  Massive slowdown in the malloc banchmark
and in LLVM/clang is very bad - we don't know what other userspace will
be so affected.

So I think we revert until this is fully understood.