Re: Linux 4.13: Reported regressions as of Sunday, 2017-09-03

[Linus Torvalds]

On Sun, Sep 3, 2017 at 2:36 AM, Thorsten Leemhuis
<regressions@xxxxxxxxxxxxx> wrote:
>
> [x86/mm/gup] e585513b76: will-it-scale.per_thread_ops -6.9% regression
> Status: Asked on the list, but looks like issue gets ignored by everyone
> Note: I'm a bit unsure if adding this issue to this list was a good
> idea. Side note: Was reported against linux-next in May already
> Reported: 2017-07-10
> http://lkml.kernel.org/r/20170710024020.GA26389@yexl-desktop
> Cause: https://git.kernel.org/torvalds/c/e585513b76f7

Sadly, while I love the concept of performance tracking, the
"will-it-scale" reports haven't really been reliable enough to really
be useful. There is a _ton_ of noise in the numbers, and the
test-cases don't seem to be stable enough to really track sanely.

I wish it was otherwise, because we also got a report of "57.3%
improvement of will-it-scale.per_process_ops" this release.

So I find the kernel test robot performance tracking very interesting
in theory, but as things stand now I think it's just that:
"interesting". Not quite ready for action.

                 Linus