From: Huang Ying <[email protected]>
In shrink_node(), to determine whether to enable cache trim mode, the
LRU size is gotten via lruvec_page_state(). That gets the value from
a per-CPU counter (mem_cgroup_per_node->lruvec_stat[]). The error of
the per-CPU counter from CPU local counting and the descendant memory
cgroups may cause some issues. We run into this in 0-Day performance
test.
0-Day uses the RAM file system as root file system, so the number of
the reclaimable file pages is very small. In the swap testing, the
inactive file LRU list will become almost empty soon. But the size of
the inactive file LRU list gotten from the per-CPU counter may keep a
much larger value (say, 33, 50, etc.). This will enable cache trim
mode, but nothing can be scanned in fact. The following pattern
repeats for long time in the test,
priority inactive_file_size cache_trim_mode
12 33 0
11 33 0
...
6 33 0
5 33 1
...
1 33 1
That is, the cache_trim_mode will be enabled wrongly when the scan
priority decreases to 5. And the problem will not be recovered for
long time.
It's hard to get the more accurate size of the inactive file list
without much more overhead. And it's hard to estimate the error of
the per-CPU counter too, because there may be many descendant memory
cgroups. But after the actual scanning, if nothing can be scanned
with the cache trim mode, it should be wrong to enable the cache trim
mode. So we can retry with the cache trim mode disabled. This patch
implement this policy.
The test results for pmbench with normal access address distribution
and 2 NVMe disks as swap on a Intel server for the base and patched
kernel are as follows.
base change patched
---------- ------ ---------
122100000 18.6% 144800000 pmbench.throughput.aps
164078 -56.3% 71686 vmstat.swap.si
166511 -20.8% 131957 vmstat.swap.so
1992 -52.2% 953.17 proc-vmstat.kswapd_high_wmark_hit_quickly
230.20 -63.8% 83.33 proc-vmstat.kswapd_low_wmark_hit_quickly
2318 -41.4% 1358 proc-vmstat.pageoutrun
From the above table, with the patch, the page reclaiming algorithm
runs more smoothly, the hot/cold pages distinguishing works better, so
the swap in/out throughput decreases considerably, and the benchmark
throughput increases 18.6%.
Signed-off-by: "Huang, Ying" <[email protected]>
Cc: Mel Gorman <[email protected]>
Cc: Johannes Weiner <[email protected]>
Cc: Vladimir Davydov <[email protected]>
Cc: Michal Hocko <[email protected]>
Cc: Joonsoo Kim <[email protected]>
---
mm/vmscan.c | 19 +++++++++++++++----
1 file changed, 15 insertions(+), 4 deletions(-)
diff --git a/mm/vmscan.c b/mm/vmscan.c
index fea6b43bc1f9..1304e3c98a14 100644
--- a/mm/vmscan.c
+++ b/mm/vmscan.c
@@ -2606,7 +2606,8 @@ static inline bool should_continue_reclaim(struct pglist_data *pgdat,
return inactive_lru_pages > pages_for_compaction;
}
-static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc)
+static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc,
+ bool skip_slab)
{
struct mem_cgroup *target_memcg = sc->target_mem_cgroup;
struct mem_cgroup *memcg;
@@ -2652,8 +2653,9 @@ static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc)
shrink_lruvec(lruvec, sc);
- shrink_slab(sc->gfp_mask, pgdat->node_id, memcg,
- sc->priority);
+ if (!skip_slab)
+ shrink_slab(sc->gfp_mask, pgdat->node_id, memcg,
+ sc->priority);
/* Record the group's reclaim efficiency */
vmpressure(sc->gfp_mask, memcg, false,
@@ -2669,6 +2671,7 @@ static void shrink_node(pg_data_t *pgdat, struct scan_control *sc)
unsigned long nr_reclaimed, nr_scanned;
struct lruvec *target_lruvec;
bool reclaimable = false;
+ bool skip_slab;
unsigned long file;
target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat);
@@ -2767,7 +2770,15 @@ static void shrink_node(pg_data_t *pgdat, struct scan_control *sc)
anon >> sc->priority;
}
- shrink_node_memcgs(pgdat, sc);
+ skip_slab = false;
+retry:
+ shrink_node_memcgs(pgdat, sc, skip_slab);
+ /* Nothing can be scanned with cache trim mode, retry without it */
+ if (sc->cache_trim_mode && sc->nr_scanned == nr_scanned) {
+ sc->cache_trim_mode = 0;
+ skip_slab = true;
+ goto retry;
+ }
if (reclaim_state) {
sc->nr_reclaimed += reclaim_state->reclaimed_slab;
--
2.30.1
On Wed, Mar 10, 2021 at 4:47 PM Huang, Ying <[email protected]> wrote:
>
> From: Huang Ying <[email protected]>
>
> In shrink_node(), to determine whether to enable cache trim mode, the
> LRU size is gotten via lruvec_page_state(). That gets the value from
> a per-CPU counter (mem_cgroup_per_node->lruvec_stat[]). The error of
> the per-CPU counter from CPU local counting and the descendant memory
> cgroups may cause some issues. We run into this in 0-Day performance
> test.
>
> 0-Day uses the RAM file system as root file system, so the number of
> the reclaimable file pages is very small. In the swap testing, the
> inactive file LRU list will become almost empty soon. But the size of
> the inactive file LRU list gotten from the per-CPU counter may keep a
> much larger value (say, 33, 50, etc.). This will enable cache trim
> mode, but nothing can be scanned in fact. The following pattern
> repeats for long time in the test,
>
> priority inactive_file_size cache_trim_mode
> 12 33 0
> 11 33 0
> ...
> 6 33 0
> 5 33 1
> ...
> 1 33 1
>
> That is, the cache_trim_mode will be enabled wrongly when the scan
> priority decreases to 5. And the problem will not be recovered for
> long time.
>
> It's hard to get the more accurate size of the inactive file list
> without much more overhead. And it's hard to estimate the error of
> the per-CPU counter too, because there may be many descendant memory
> cgroups. But after the actual scanning, if nothing can be scanned
> with the cache trim mode, it should be wrong to enable the cache trim
> mode. So we can retry with the cache trim mode disabled. This patch
> implement this policy.
Instead of playing with the already complicated heuristics, we should
improve the accuracy of the lruvec stats. Johannes already fixed the
memcg stats using rstat infrastructure and Tejun has suggestions on
how to use rstat infrastructure efficiently for lruvec stats at
https://lore.kernel.org/linux-mm/[email protected]/.
Hi, Butt,
Shakeel Butt <[email protected]> writes:
> On Wed, Mar 10, 2021 at 4:47 PM Huang, Ying <[email protected]> wrote:
>>
>> From: Huang Ying <[email protected]>
>>
>> In shrink_node(), to determine whether to enable cache trim mode, the
>> LRU size is gotten via lruvec_page_state(). That gets the value from
>> a per-CPU counter (mem_cgroup_per_node->lruvec_stat[]). The error of
>> the per-CPU counter from CPU local counting and the descendant memory
>> cgroups may cause some issues. We run into this in 0-Day performance
>> test.
>>
>> 0-Day uses the RAM file system as root file system, so the number of
>> the reclaimable file pages is very small. In the swap testing, the
>> inactive file LRU list will become almost empty soon. But the size of
>> the inactive file LRU list gotten from the per-CPU counter may keep a
>> much larger value (say, 33, 50, etc.). This will enable cache trim
>> mode, but nothing can be scanned in fact. The following pattern
>> repeats for long time in the test,
>>
>> priority inactive_file_size cache_trim_mode
>> 12 33 0
>> 11 33 0
>> ...
>> 6 33 0
>> 5 33 1
>> ...
>> 1 33 1
>>
>> That is, the cache_trim_mode will be enabled wrongly when the scan
>> priority decreases to 5. And the problem will not be recovered for
>> long time.
>>
>> It's hard to get the more accurate size of the inactive file list
>> without much more overhead. And it's hard to estimate the error of
>> the per-CPU counter too, because there may be many descendant memory
>> cgroups. But after the actual scanning, if nothing can be scanned
>> with the cache trim mode, it should be wrong to enable the cache trim
>> mode. So we can retry with the cache trim mode disabled. This patch
>> implement this policy.
>
> Instead of playing with the already complicated heuristics, we should
> improve the accuracy of the lruvec stats. Johannes already fixed the
> memcg stats using rstat infrastructure and Tejun has suggestions on
> how to use rstat infrastructure efficiently for lruvec stats at
> https://lore.kernel.org/linux-mm/[email protected]/.
Thanks for your information! It should be better if we can improve the
accuracy of lruvec stats without much overhead. But that may be not a
easy task.
If my understanding were correct, what Tejun suggested is to add a fast
read interface to rstat to be used in hot path. And its accuracy is
similar as that of traditional per-CPU counter. But if we can regularly
update the lruvec rstat with something like vmstat_update(), that should
be OK for the issue described in this patch.
Best Regards,
Huang, Ying
On Thu, Mar 11, 2021 at 12:52 AM Huang, Ying <[email protected]> wrote:
>
> Hi, Butt,
>
> Shakeel Butt <[email protected]> writes:
>
> > On Wed, Mar 10, 2021 at 4:47 PM Huang, Ying <[email protected]> wrote:
> >>
> >> From: Huang Ying <[email protected]>
> >>
> >> In shrink_node(), to determine whether to enable cache trim mode, the
> >> LRU size is gotten via lruvec_page_state(). That gets the value from
> >> a per-CPU counter (mem_cgroup_per_node->lruvec_stat[]). The error of
> >> the per-CPU counter from CPU local counting and the descendant memory
> >> cgroups may cause some issues. We run into this in 0-Day performance
> >> test.
> >>
> >> 0-Day uses the RAM file system as root file system, so the number of
> >> the reclaimable file pages is very small. In the swap testing, the
> >> inactive file LRU list will become almost empty soon. But the size of
> >> the inactive file LRU list gotten from the per-CPU counter may keep a
> >> much larger value (say, 33, 50, etc.). This will enable cache trim
> >> mode, but nothing can be scanned in fact. The following pattern
> >> repeats for long time in the test,
> >>
> >> priority inactive_file_size cache_trim_mode
> >> 12 33 0
> >> 11 33 0
> >> ...
> >> 6 33 0
> >> 5 33 1
> >> ...
> >> 1 33 1
> >>
> >> That is, the cache_trim_mode will be enabled wrongly when the scan
> >> priority decreases to 5. And the problem will not be recovered for
> >> long time.
> >>
> >> It's hard to get the more accurate size of the inactive file list
> >> without much more overhead. And it's hard to estimate the error of
> >> the per-CPU counter too, because there may be many descendant memory
> >> cgroups. But after the actual scanning, if nothing can be scanned
> >> with the cache trim mode, it should be wrong to enable the cache trim
> >> mode. So we can retry with the cache trim mode disabled. This patch
> >> implement this policy.
> >
> > Instead of playing with the already complicated heuristics, we should
> > improve the accuracy of the lruvec stats. Johannes already fixed the
> > memcg stats using rstat infrastructure and Tejun has suggestions on
> > how to use rstat infrastructure efficiently for lruvec stats at
> > https://lore.kernel.org/linux-mm/[email protected]/.
>
> Thanks for your information! It should be better if we can improve the
> accuracy of lruvec stats without much overhead. But that may be not a
> easy task.
>
> If my understanding were correct, what Tejun suggested is to add a fast
> read interface to rstat to be used in hot path. And its accuracy is
> similar as that of traditional per-CPU counter. But if we can regularly
> update the lruvec rstat with something like vmstat_update(), that should
> be OK for the issue described in this patch.
>
This is also my understanding. Tejun, please correct us if we misunderstood you.
BTW Johannes was working on rstat-based lruvec stats patch. Johannes,
are you planning to work on the optimization Tejun has suggested.
Hello,
On Sun, Mar 14, 2021 at 01:58:33PM -0700, Shakeel Butt wrote:
> > If my understanding were correct, what Tejun suggested is to add a fast
> > read interface to rstat to be used in hot path. And its accuracy is
> > similar as that of traditional per-CPU counter. But if we can regularly
> > update the lruvec rstat with something like vmstat_update(), that should
> > be OK for the issue described in this patch.
> >
>
> This is also my understanding. Tejun, please correct us if we misunderstood you.
Yeah, that was what I had on mind. Instead of always flushing on read, have
a variant where flushing is explicit and trigger that periodically (or
whichever way appropriate).
Thanks.
--
tejun