On Tue 27-11-18 19:17:27, Michal Hocko wrote:
> On Tue 27-11-18 09:08:50, Linus Torvalds wrote:
> > On Mon, Nov 26, 2018 at 10:24 PM kernel test robot
> > <rong.a.c...@intel.com> wrote:
> > >
> > > FYI, we noticed a -61.3% regression of vm-scalability.throughput due
> > > to commit ac5b2c18911f ("mm: thp: relax __GFP_THISNODE for
> > > MADV_HUGEPAGE mappings")
> > 
> > Well, that's certainly noticeable and not good.
> > 
> > Andrea, I suspect it might be causing fights with auto numa migration..
> > 
> > Lots more system time, but also look at this:
> > 
> > >    1122389 ±  9%     +17.2%    1315380 ±  4%  proc-vmstat.numa_hit
> > >     214722 ±  5%     +21.6%     261076 ±  3%  
> > > proc-vmstat.numa_huge_pte_updates
> > >    1108142 ±  9%     +17.4%    1300857 ±  4%  proc-vmstat.numa_local
> > >     145368 ± 48%     +63.1%     237050 ± 17%  proc-vmstat.numa_miss
> > >     159615 ± 44%     +57.6%     251573 ± 16%  proc-vmstat.numa_other
> > >     185.50 ± 81%   +8278.6%      15542 ± 40%  
> > > proc-vmstat.numa_pages_migrated
> > 
> > Should the commit be reverted? Or perhaps at least modified?
> 
> Well, the commit is trying to revert to the behavior before
> 5265047ac301 because there are real usecases that suffered from that
> change and bug reports as a result of that.
> 
> will-it-scale is certainly worth considering but it is an artificial
> testcase. A higher NUMA miss rate is an expected side effect of the
> patch because the fallback to a different NUMA node is more likely. The
> __GFP_THISNODE side effect is basically introducing node-reclaim
> behavior for THPages. Another thing is that there is no good behavior
> for everybody. Reclaim locally vs. THP on a remote node is hard to
> tell by default. We have discussed that at length and there were some
> conclusions. One of them is that we need a numa policy to tell whether
> a expensive localility is preferred over remote allocation.  Also we
> definitely need a better pro-active defragmentation to allow larger
> pages on a local node. This is a work in progress and this patch is a
> stop gap fix.

Btw. the associated discussion is 
http://lkml.kernel.org/r/20180925120326.24392-1-mho...@kernel.org

-- 
Michal Hocko
SUSE Labs

Reply via email to