On Wed, Jan 31, 2024 at 10:01:27AM -0800, T.J. Mercier wrote:
> On Wed, Jan 31, 2024 at 9:51 AM Johannes Weiner <hannes@xxxxxxxxxxx> wrote:
> >
> > On Wed, Jan 31, 2024 at 04:24:41PM +0000, T.J. Mercier wrote:
> > > Before 388536ac291 ("mm:vmscan: fix inaccurate reclaim during proactive
> > > reclaim") we passed the number of pages for the reclaim request directly
> > > to try_to_free_mem_cgroup_pages, which could lead to significant
> > > overreclaim in order to achieve fairness. After 0388536ac291 the number
> > > of pages was limited to a maxmimum of 32 (SWAP_CLUSTER_MAX) to reduce
> > > the amount of overreclaim. However such a small chunk size caused a
> > > regression in reclaim performance due to many more reclaim start/stop
> > > cycles inside memory_reclaim.
> > >
> > > Instead of limiting reclaim chunk size to the SWAP_CLUSTER_MAX constant,
> > > adjust the chunk size proportionally with number of pages left to
> > > reclaim. This allows for higher reclaim efficiency with large chunk
> > > sizes during the beginning of memory_reclaim, and reduces the amount of
> > > potential overreclaim by using small chunk sizes as the total reclaim
> > > amount is approached. Using 1/4 of the amount left to reclaim as the
> > > chunk size gives a good compromise between reclaim performance and
> > > overreclaim:
> > >
> > > root - full reclaim pages/sec time (sec)
> > > pre-0388536ac291 : 68047 10.46
> > > post-0388536ac291 : 13742 inf
> > > (reclaim-reclaimed)/4 : 67352 10.51
> > >
> > > /uid_0 - 1G reclaim pages/sec time (sec) overreclaim (MiB)
> > > pre-0388536ac291 : 258822 1.12 107.8
> > > post-0388536ac291 : 105174 2.49 3.5
> > > (reclaim-reclaimed)/4 : 233396 1.12 -7.4
> > >
> > > /uid_0 - full reclaim pages/sec time (sec)
> > > pre-0388536ac291 : 72334 7.09
> > > post-0388536ac291 : 38105 14.45
> > > (reclaim-reclaimed)/4 : 72914 6.96
> > >
> > > Fixes: 0388536ac291 ("mm:vmscan: fix inaccurate reclaim during proactive reclaim")
> > > Signed-off-by: T.J. Mercier <tjmercier@xxxxxxxxxx>
> > > ---
> > > mm/memcontrol.c | 3 ++-
> > > 1 file changed, 2 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> > > index 46d8d02114cf..d68fb89eadd2 100644
> > > --- a/mm/memcontrol.c
> > > +++ b/mm/memcontrol.c
> > > @@ -6977,7 +6977,8 @@ static ssize_t memory_reclaim(struct kernfs_open_file *of, char *buf,
> > > lru_add_drain_all();
> > >
> > > reclaimed = try_to_free_mem_cgroup_pages(memcg,
> > > - min(nr_to_reclaim - nr_reclaimed, SWAP_CLUSTER_MAX),
> > > + max((nr_to_reclaim - nr_reclaimed) / 4,
> > > + (nr_to_reclaim - nr_reclaimed) % 4),
> >
> > I don't see why the % 4 is needed. It only kicks in when the delta
> > drops below 4, but try_to_free_mem_cgroup_pages() already has
> >
> > .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
> >
> > so it looks like dead code.
>
> That right, it's only there for when the integer division reaches
> zero. I didn't want to assume anything about the implementation of
> try_to_free_mem_cgroup_pages, but I can just remove it entirely if
> you'd like.
What do others think?
We rely on the rounding up in a few other places and it's been doing
that for a decade. Maybe lampshade it for the benefit of the reader:
/* Will converge on zero, but reclaim enforces a minimum */
but otherwise there is IMO no need to have defensive extra code.
