Re: [PATCH] mm: Do not deactivate when the cgroup has plenty inactive page

[Michal Hocko <mhocko@suse.com>]

On Wed 14-10-20 22:21:58, zhong jiang wrote:
> 
> On 2020/9/30 12:02 上午, Michal Hocko wrote:
> > On Sun 27-09-20 10:39:22, zhong jiang wrote:
> > > On 2020/9/25 8:07 下午, Michal Hocko wrote:
> > > > On Fri 25-09-20 19:49:12, zhongjiang-ali wrote:
> > > > > After appling the series patches(mm: fix page aging across multiple cgroups),
> > > > > cgroup memory reclaim strategy is based on reclaim root's inactive:active
> > > > > ratio. if the target lruvec need to deactivate, its children cgroup also will
> > > > > deactivate. That will result in hot page to be reclaimed and other cgroup's
> > > > > cold page will be left, which is not expected.
> > > > > 
> > > > > The patch will not force deactivate when inactive_is_low is not true unless
> > > > > we has scanned the inactive list and memory is unable to reclaim.
> > > > Do you have any data to present?
> > > I write an testcase that cgroup B has a lot of hot pagecache and cgroup C
> > > is full of cold pagecache.  and
> > > 
> > > their parent cgroup A will trigger the reclaim due of it's limit has been
> > > breached.
> > > 
> > > 
> > > The testcase should assume that we should not reclaim the  hot pagecache in
> > > cgroup B because C has
> > > 
> > > plenty cold pagecache.   Unfortunately,  I can see cgroup B hot pagecache
> > > has been decreased when
> > > 
> > > cgroup A trigger the reclaim.
> > Thank you, this is more or less what've expected from your initial
> > description. An extended form would be preferable for the changelog to
> > make the setup more clear. But you still haven't quantified the effect.
> > It would be also good to describe what is the effect on the workload
> > described by 53138cea7f39 ("mm: vmscan: move file exhaustion detection
> > to the node level"). A more extended description on the implementation
> > would be nice as well.
> 
> Hi,  Michal
> 
> I'm sorry for lately reply due of a long vacation.  But that indeed breach
> the initial purpose.
> 
> Do you think the patch make sense, or any benchmark can be recommended to
> get some data.

To be honest I have really hard time to grasp what is the effect of this
patch. Also memory reclaim tuning without any data doesn't really sound
convincing. Do you see any real workload that is benefiting from this
change or this is mostly a based on reading the code and a theoretical
concern?

Please understand that the existing balancing is quite complex and any
changes should be carfully analyzed and described.

-- 
Michal Hocko
SUSE Labs