Re: [PATCH] mm: fix draining remote pageset

[Michal Hocko <mhocko@suse.com>]

On Tue 22-08-23 06:31:42, Huang, Ying wrote:
> Michal Hocko <mhocko@suse.com> writes:
> 
> > On Mon 21-08-23 16:30:18, Huang, Ying wrote:
> >> Michal Hocko <mhocko@suse.com> writes:
> >> 
> >> > On Wed 16-08-23 15:08:23, Huang, Ying wrote:
> >> >> Michal Hocko <mhocko@suse.com> writes:
> >> >> 
> >> >> > On Mon 14-08-23 09:59:51, Huang, Ying wrote:
> >> >> >> Hi, Michal,
> >> >> >> 
> >> >> >> Michal Hocko <mhocko@suse.com> writes:
> >> >> >> 
> >> >> >> > On Fri 11-08-23 17:08:19, Huang Ying wrote:
> >> >> >> >> If there is no memory allocation/freeing in the remote pageset after
> >> >> >> >> some time (3 seconds for now), the remote pageset will be drained to
> >> >> >> >> avoid memory wastage.
> >> >> >> >> 
> >> >> >> >> But in the current implementation, vmstat updater worker may not be
> >> >> >> >> re-queued when we are waiting for the timeout (pcp->expire != 0) if
> >> >> >> >> there are no vmstat changes, for example, when CPU goes idle.
> >> >> >> >
> >> >> >> > Why is that a problem?
> >> >> >> 
> >> >> >> The pages of the remote zone may be kept in the local per-CPU pageset
> >> >> >> for long time as long as there's no page allocation/freeing on the
> >> >> >> logical CPU.  In addition to the logical CPU goes idle, this is also
> >> >> >> possible if the logical CPU is busy in the user space.
> >> >> >
> >> >> > But why is this a problem? Is the scale of the problem sufficient to
> >> >> > trigger out of memory situations or be otherwise harmful?
> >> >> 
> >> >> This may trigger premature page reclaiming.  The pages in the PCP of the
> >> >> remote zone would have been freed to satisfy the page allocation for the
> >> >> remote zone to avoid page reclaiming.  It's highly possible that the
> >> >> local CPU just allocate/free from/to the remote zone temporarily.
> >> >
> >> > I am slightly confused here but I suspect by zone you mean remote pcp.
> >> > But more importantly is this a concern seen in real workload? Can you
> >> > quantify it in some manner? E.g. with this patch we have X more kswapd
> >> > scanning or even hit direct reclaim much less often.
> >> >> So,
> >> >> we should free PCP pages of the remote zone if there is no page
> >> >> allocation/freeing from/to the remote zone for 3 seconds.
> >> >
> >> > Well, I would argue this depends a lot. There are workloads which really
> >> > like to have CPUs idle and yet they would like to benefit from the
> >> > allocator fast path after that CPU goes out of idle because idling is
> >> > their power saving opportunity while workloads want to act quickly after
> >> > there is something to run.
> >> >
> >> > That being said, we really need some numbers (ideally from real world)
> >> > that proves this is not just a theoretical concern.
> >> 
> >> The behavior to drain the PCP of the remote zone (that is, remote PCP)
> >> was introduced in commit 4ae7c03943fc ("[PATCH] Periodically drain non
> >> local pagesets").  The goal of draining was well documented in the
> >> change log.  IIUC, some of your questions can be answered there?
> >> 
> >> This patch just restores the original behavior changed by commit
> >> 7cc36bbddde5 ("vmstat: on-demand vmstat workers V8").
> >
> > Let me repeat. You need some numbers to show this is needed.
> 
> I have done some test for this patch as follows,
> 
> - Run some workloads, use `numactl` to bind CPU to node 0 and memory to
>   node 1.  So the PCP of the CPU on node 0 for zone on node 1 will be
>   filled.
> 
> - After workloads finish, idle for 60s
> 
> - Check /proc/zoneinfo
> 
> With the original kernel, the number of pages in the PCP of the CPU on
> node 0 for zone on node 1 is non-zero after idle.  With the patched
> kernel, that becomes 0 after idle.  We avoid to keep pages in the remote
> PCP during idle.
> 
> This is the number I have.  If you think it isn't enough to justify the
> patch, then I'm OK too (although I think it's enough).  Because the
> remote PCP will be drained later when some pages are allocated/freed on
> the CPU.

Yes, this doesn't really show any actual correctness problem so I do not
think this is sufficient to change the code. You would need to show that
the existing behavior is actively harmful.

-- 
Michal Hocko
SUSE Labs