On Tue, Jul 23, 2024 at 11:42:33AM GMT, Oliver Sang wrote:
> hi, Roman,
>
> On Sat, Jul 20, 2024 at 01:14:16AM +0800, Roman Gushchin wrote:
> > On Tue, Jul 16, 2024 at 03:53:25PM +0800, Oliver Sang wrote:
> > > hi, Roman,
> > >
> > > On Mon, Jul 15, 2024 at 10:18:39PM +0000, Roman Gushchin wrote:
> > > > On Mon, Jul 15, 2024 at 10:14:31PM +0800, Oliver Sang wrote:
> > > > > hi, Roman Gushchin,
> > > > >
> > > > > On Fri, Jul 12, 2024 at 07:03:31PM +0000, Roman Gushchin wrote:
> > > > > > On Fri, Jul 12, 2024 at 02:04:48PM +0800, kernel test robot wrote:
> > > > > > >
> > > > > > >
> > > > > > > Hello,
> > > > > > >
> > > > > > > kernel test robot noticed a -29.4% regression of aim7.jobs-per-min on:
> > > > > > >
> > > > > > >
> > > > > > > commit: 98c9daf5ae6be008f78c07b744bcff7bcc6e98da ("mm: memcg: guard memcg1-specific members of struct mem_cgroup_per_node")
> > > > > > > https://git.kernel.org/cgit/linux/kernel/git/next/linux-next.git master
> > > > > >
> > > > > > Hello,
> > > > > >
> > > > > > thank you for the report!
> > > > > >
> > > > > > I'd expect that the regression should be fixed by the commit
> > > > > > "mm: memcg: add cache line padding to mem_cgroup_per_node".
> > > > > >
> > > > > > Can you, please, confirm that it's not the case?
> > > > > >
> > > > > > Thank you!
> > > > >
> > > > > in our this aim7 test, we found the performance partially recovered by
> > > > > "mm: memcg: add cache line padding to mem_cgroup_per_node" but not fully
> > > >
> > > > Thank you for providing the detailed information!
> > > >
> > > > Can you, please, check if the following patch resolves the regression entirely?
> > >
> > > no. in our tests, the following patch has little impact.
> > > I directly apply it upon 6df13230b6 (if this is not the proper applyment, please
> > > let me know, thanks)
> >
> > Hm, interesting. And thank you for the confirmation, you did everything correct.
> > Because the only thing the original patch did was a removal of few fields from
> > the mem_cgroup_per_node struct, there are not many options left here.
> > Would you mind to try the following patch?
> >
> > Thank you and really appreciate your help!
>
> you are welcome!
>
> though we saw there are further discussions, we still share our test results to
> you.
>
> in our tests, by your new version patch, the regression is entirely resoloved.
>
[...]
This is very interesting as this shows (possibly) there was false
sharing between lruvec and lru_zone_size. I will check if these two
fields were accidentally on different cacheline before the series.
> >
> >
> > diff --git a/include/linux/memcontrol.h b/include/linux/memcontrol.h
> > index 7e2eb091049a..0e5bf25d324f 100644
> > --- a/include/linux/memcontrol.h
> > +++ b/include/linux/memcontrol.h
> > @@ -109,6 +109,7 @@ struct mem_cgroup_per_node {
> >
> > /* Fields which get updated often at the end. */
> > struct lruvec lruvec;
> > + CACHELINE_PADDING(_pad2_);
> > unsigned long lru_zone_size[MAX_NR_ZONES][NR_LRU_LISTS];
> > struct mem_cgroup_reclaim_iter iter;
> > };
> >
> >
> >
