On Fri 08-02-19 10:55:07, Jan Kara wrote: > On Thu 07-02-19 21:37:27, Andrew Morton wrote: > > On Thu, 7 Feb 2019 11:27:50 +0100 Jan Kara <jack@xxxxxxx> wrote: > > > > > On Fri 01-02-19 09:19:04, Dave Chinner wrote: > > > > Maybe for memcgs, but that's exactly the oppose of what we want to > > > > do for global caches (e.g. filesystem metadata caches). We need to > > > > make sure that a single, heavily pressured cache doesn't evict small > > > > caches that lower pressure but are equally important for > > > > performance. > > > > > > > > e.g. I've noticed recently a significant increase in RMW cycles in > > > > XFS inode cache writeback during various benchmarks. It hasn't > > > > affected performance because the machine has IO and CPU to burn, but > > > > on slower machines and storage, it will have a major impact. > > > > > > Just as a data point, our performance testing infrastructure has bisected > > > down to the commits discussed in this thread as the cause of about 40% > > > regression in XFS file delete performance in bonnie++ benchmark. > > > > > > > Has anyone done significant testing with Rik's maybe-fix? > > I will give it a spin with bonnie++ today. We'll see what comes out. OK, I did a bonnie++ run with Rik's patch (on top of 4.20 to rule out other differences). This machine does not show so big differences in bonnie++ numbers but the difference is still clearly visible. The results are (averages of 5 runs): Revert Base Rik SeqCreate del 78.04 ( 0.00%) 98.18 ( -25.81%) 90.90 ( -16.48%) RandCreate del 87.68 ( 0.00%) 95.01 ( -8.36%) 87.66 ( 0.03%) 'Revert' is 4.20 with "mm: don't reclaim inodes with many attached pages" and "mm: slowly shrink slabs with a relatively small number of objects" reverted. 'Base' is the kernel without any reverts. 'Rik' is a 4.20 with Rik's patch applied. The numbers are time to do a batch of deletes so lower is better. You can see that the patch did help somewhat but it was not enough to close the gap when files are deleted in 'readdir' order. Honza > > From: Rik van Riel <riel@xxxxxxxxxxx> > > Subject: mm, slab, vmscan: accumulate gradual pressure on small slabs > > > > There are a few issues with the way the number of slab objects to scan is > > calculated in do_shrink_slab. First, for zero-seek slabs, we could leave > > the last object around forever. That could result in pinning a dying > > cgroup into memory, instead of reclaiming it. The fix for that is > > trivial. > > > > Secondly, small slabs receive much more pressure, relative to their size, > > than larger slabs, due to "rounding up" the minimum number of scanned > > objects to batch_size. > > > > We can keep the pressure on all slabs equal relative to their size by > > accumulating the scan pressure on small slabs over time, resulting in > > sometimes scanning an object, instead of always scanning several. > > > > This results in lower system CPU use, and a lower major fault rate, as > > actively used entries from smaller caches get reclaimed less aggressively, > > and need to be reloaded/recreated less often. > > > > [akpm@xxxxxxxxxxxxxxxxxxxx: whitespace fixes, per Roman] > > [riel@xxxxxxxxxxx: couple of fixes] > > Link: http://lkml.kernel.org/r/20190129142831.6a373403@xxxxxxxxxxxxxxxxxxxx > > Link: http://lkml.kernel.org/r/20190128143535.7767c397@xxxxxxxxxxxxxxxxxxxx > > Fixes: 4b85afbdacd2 ("mm: zero-seek shrinkers") > > Fixes: 172b06c32b94 ("mm: slowly shrink slabs with a relatively small number of objects") > > Signed-off-by: Rik van Riel <riel@xxxxxxxxxxx> > > Tested-by: Chris Mason <clm@xxxxxx> > > Acked-by: Roman Gushchin <guro@xxxxxx> > > Acked-by: Johannes Weiner <hannes@xxxxxxxxxxx> > > Cc: Dave Chinner <dchinner@xxxxxxxxxx> > > Cc: Jonathan Lemon <bsd@xxxxxx> > > Cc: Jan Kara <jack@xxxxxxx> > > Cc: <stable@xxxxxxxxxxxxxxx> > > > > Signed-off-by: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> > > --- > > > > > > --- a/include/linux/shrinker.h~mmslabvmscan-accumulate-gradual-pressure-on-small-slabs > > +++ a/include/linux/shrinker.h > > @@ -65,6 +65,7 @@ struct shrinker { > > > > long batch; /* reclaim batch size, 0 = default */ > > int seeks; /* seeks to recreate an obj */ > > + int small_scan; /* accumulate pressure on slabs with few objects */ > > unsigned flags; > > > > /* These are for internal use */ > > --- a/mm/vmscan.c~mmslabvmscan-accumulate-gradual-pressure-on-small-slabs > > +++ a/mm/vmscan.c > > @@ -488,18 +488,30 @@ static unsigned long do_shrink_slab(stru > > * them aggressively under memory pressure to keep > > * them from causing refetches in the IO caches. > > */ > > - delta = freeable / 2; > > + delta = (freeable + 1) / 2; > > } > > > > /* > > * Make sure we apply some minimal pressure on default priority > > - * even on small cgroups. Stale objects are not only consuming memory > > + * even on small cgroups, by accumulating pressure across multiple > > + * slab shrinker runs. Stale objects are not only consuming memory > > * by themselves, but can also hold a reference to a dying cgroup, > > * preventing it from being reclaimed. A dying cgroup with all > > * corresponding structures like per-cpu stats and kmem caches > > * can be really big, so it may lead to a significant waste of memory. > > */ > > - delta = max_t(unsigned long long, delta, min(freeable, batch_size)); > > + if (!delta && shrinker->seeks) { > > + unsigned long nr_considered; > > + > > + shrinker->small_scan += freeable; > > + nr_considered = shrinker->small_scan >> priority; > > + > > + delta = 4 * nr_considered; > > + do_div(delta, shrinker->seeks); > > + > > + if (delta) > > + shrinker->small_scan -= nr_considered << priority; > > + } > > > > total_scan += delta; > > if (total_scan < 0) { > > _ > > > -- > Jan Kara <jack@xxxxxxxx> > SUSE Labs, CR -- Jan Kara <jack@xxxxxxxx> SUSE Labs, CR
