On Mon, Oct 23, 2023 at 11:25 AM Yosry Ahmed <yosryahmed@xxxxxxxxxx> wrote:
>
> On Sun, Oct 22, 2023 at 6:34 PM Feng Tang <feng.tang@xxxxxxxxx> wrote:
> >
> > On Sat, Oct 21, 2023 at 01:42:58AM +0800, Yosry Ahmed wrote:
> > > On Fri, Oct 20, 2023 at 10:23 AM Shakeel Butt <shakeelb@xxxxxxxxxx> wrote:
> > > >
> > > > On Fri, Oct 20, 2023 at 9:18 AM kernel test robot <oliver.sang@xxxxxxxxx> wrote:
> > > > >
> > > > >
> > > > >
> > > > > Hello,
> > > > >
> > > > > kernel test robot noticed a -25.8% regression of will-it-scale.per_thread_ops on:
> > > > >
> > > > >
> > > > > commit: 51d74c18a9c61e7ee33bc90b522dd7f6e5b80bb5 ("[PATCH v2 3/5] mm: memcg: make stats flushing threshold per-memcg")
> > > > > url: https://github.com/intel-lab-lkp/linux/commits/Yosry-Ahmed/mm-memcg-change-flush_next_time-to-flush_last_time/20231010-112257
> > > > > base: https://git.kernel.org/cgit/linux/kernel/git/akpm/mm.git mm-everything
> > > > > patch link: https://lore.kernel.org/all/20231010032117.1577496-4-yosryahmed@xxxxxxxxxx/
> > > > > patch subject: [PATCH v2 3/5] mm: memcg: make stats flushing threshold per-memcg
> > > > >
> > > > > testcase: will-it-scale
> > > > > test machine: 104 threads 2 sockets (Skylake) with 192G memory
> > > > > parameters:
> > > > >
> > > > > nr_task: 100%
> > > > > mode: thread
> > > > > test: fallocate1
> > > > > cpufreq_governor: performance
> > > > >
> > > > >
> > > > > In addition to that, the commit also has significant impact on the following tests:
> > > > >
> > > > > +------------------+---------------------------------------------------------------+
> > > > > | testcase: change | will-it-scale: will-it-scale.per_thread_ops -30.0% regression |
> > > > > | test machine | 104 threads 2 sockets (Skylake) with 192G memory |
> > > > > | test parameters | cpufreq_governor=performance |
> > > > > | | mode=thread |
> > > > > | | nr_task=50% |
> > > > > | | test=fallocate1 |
> > > > > +------------------+---------------------------------------------------------------+
> > > > >
> > > >
> > > > Yosry, I don't think 25% to 30% regression can be ignored. Unless
> > > > there is a quick fix, IMO this series should be skipped for the
> > > > upcoming kernel open window.
> > >
> > > I am currently looking into it. It's reasonable to skip the next merge
> > > window if a quick fix isn't found soon.
> > >
> > > I am surprised by the size of the regression given the following:
> > > 1.12 ą 5% +1.4 2.50 ą 2%
> > > perf-profile.self.cycles-pp.__mod_memcg_lruvec_state
> > >
> > > IIUC we are only spending 1% more time in __mod_memcg_lruvec_state().
> >
> > Yes, this is kind of confusing. And we have seen similar cases before,
> > espcially for micro benchmark like will-it-scale, stressng, netperf
> > etc, the change to those functions in hot path was greatly amplified
> > in the final benchmark score.
> >
> > In a netperf case, https://lore.kernel.org/lkml/20220619150456.GB34471@xsang-OptiPlex-9020/
> > the affected functions have around 10% change in perf's cpu-cycles,
> > and trigger 69% regression. IIRC, micro benchmarks are very sensitive
> > to those statistics update, like memcg's and vmstat.
> >
>
> Thanks for clarifying. I am still trying to reproduce locally but I am
> running into some quirks with tooling. I may have to run a modified
> version of the fallocate test manually. Meanwhile, I noticed that the
> patch was tested without the fixlet that I posted [1] for it,
> understandably. Would it be possible to get some numbers with that
> fixlet? It should reduce the total number of contended atomic
> operations, so it may help.
>
> [1]https://lore.kernel.org/lkml/CAJD7tkZDarDn_38ntFg5bK2fAmFdSe+Rt6DKOZA7Sgs_kERoVA@xxxxxxxxxxxxxx/
>
> I am also wondering if aligning the stats_updates atomic will help.
> Right now it may share a cacheline with some items of the
> events_pending array. The latter may be dirtied during a flush and
> unnecessarily dirty the former, but the chances are slim to be honest.
> If it's easy to test such a diff, that would be nice, but I don't
> expect a lot of difference:
>
> diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> index 7cbc7d94eb65..a35fce653262 100644
> --- a/mm/memcontrol.c
> +++ b/mm/memcontrol.c
> @@ -646,7 +646,7 @@ struct memcg_vmstats {
> unsigned long events_pending[NR_MEMCG_EVENTS];
>
> /* Stats updates since the last flush */
> - atomic64_t stats_updates;
> + atomic64_t stats_updates ____cacheline_aligned_in_smp;
> };
>
> /*
I still could not run the benchmark, but I used a version of
fallocate1.c that does 1 million iterations. I ran 100 in parallel.
This showed ~13% regression with the patch, so not the same as the
will-it-scale version, but it could be an indicator.
With that, I did not see any improvement with the fixlet above or
___cacheline_aligned_in_smp. So you can scratch that.
I did, however, see some improvement with reducing the indirection
layers by moving stats_updates directly into struct mem_cgroup. The
regression in my manual testing went down to 9%. Still not great, but
I am wondering how this reflects on the benchmark. If you're able to
test it that would be great, the diff is below. Meanwhile I am still
looking for other improvements that can be made.
diff --git a/include/linux/memcontrol.h b/include/linux/memcontrol.h
index f64ac140083e..b4dfcd8b9cc1 100644
--- a/include/linux/memcontrol.h
+++ b/include/linux/memcontrol.h
@@ -270,6 +270,9 @@ struct mem_cgroup {
CACHELINE_PADDING(_pad1_);
+ /* Stats updates since the last flush */
+ atomic64_t stats_updates;
+
/* memory.stat */
struct memcg_vmstats *vmstats;
@@ -309,6 +312,7 @@ struct mem_cgroup {
atomic_t moving_account;
struct task_struct *move_lock_task;
+ unsigned int __percpu *stats_updates_percpu;
struct memcg_vmstats_percpu __percpu *vmstats_percpu;
#ifdef CONFIG_CGROUP_WRITEBACK
diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index 7cbc7d94eb65..e5d2f3d4d874 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -627,9 +627,6 @@ struct memcg_vmstats_percpu {
/* Cgroup1: threshold notifications & softlimit tree updates */
unsigned long nr_page_events;
unsigned long targets[MEM_CGROUP_NTARGETS];
-
- /* Stats updates since the last flush */
- unsigned int stats_updates;
};
struct memcg_vmstats {
@@ -644,9 +641,6 @@ struct memcg_vmstats {
/* Pending child counts during tree propagation */
long state_pending[MEMCG_NR_STAT];
unsigned long events_pending[NR_MEMCG_EVENTS];
-
- /* Stats updates since the last flush */
- atomic64_t stats_updates;
};
/*
@@ -695,14 +689,14 @@ static void memcg_stats_unlock(void)
static bool memcg_should_flush_stats(struct mem_cgroup *memcg)
{
- return atomic64_read(&memcg->vmstats->stats_updates) >
+ return atomic64_read(&memcg->stats_updates) >
MEMCG_CHARGE_BATCH * num_online_cpus();
}
static inline void memcg_rstat_updated(struct mem_cgroup *memcg, int val)
{
int cpu = smp_processor_id();
- unsigned int x;
+ unsigned int *stats_updates_percpu;
if (!val)
return;
@@ -710,10 +704,10 @@ static inline void memcg_rstat_updated(struct
mem_cgroup *memcg, int val)
cgroup_rstat_updated(memcg->css.cgroup, cpu);
for (; memcg; memcg = parent_mem_cgroup(memcg)) {
- x = __this_cpu_add_return(memcg->vmstats_percpu->stats_updates,
- abs(val));
+ stats_updates_percpu =
this_cpu_ptr(memcg->stats_updates_percpu);
- if (x < MEMCG_CHARGE_BATCH)
+ *stats_updates_percpu += abs(val);
+ if (*stats_updates_percpu < MEMCG_CHARGE_BATCH)
continue;
/*
@@ -721,8 +715,8 @@ static inline void memcg_rstat_updated(struct
mem_cgroup *memcg, int val)
* redundant. Avoid the overhead of the atomic update.
*/
if (!memcg_should_flush_stats(memcg))
- atomic64_add(x, &memcg->vmstats->stats_updates);
- __this_cpu_write(memcg->vmstats_percpu->stats_updates, 0);
+ atomic64_add(*stats_updates_percpu,
&memcg->stats_updates);
+ *stats_updates_percpu = 0;
}
}
@@ -5467,6 +5461,7 @@ static void __mem_cgroup_free(struct mem_cgroup *memcg)
free_mem_cgroup_per_node_info(memcg, node);
kfree(memcg->vmstats);
free_percpu(memcg->vmstats_percpu);
+ free_percpu(memcg->stats_updates_percpu);
kfree(memcg);
}
@@ -5504,6 +5499,11 @@ static struct mem_cgroup *mem_cgroup_alloc(void)
if (!memcg->vmstats_percpu)
goto fail;
+ memcg->stats_updates_percpu = alloc_percpu_gfp(unsigned int,
+ GFP_KERNEL_ACCOUNT);
+ if (!memcg->stats_updates_percpu)
+ goto fail;
+
for_each_node(node)
if (alloc_mem_cgroup_per_node_info(memcg, node))
goto fail;
@@ -5735,10 +5735,12 @@ static void mem_cgroup_css_rstat_flush(struct
cgroup_subsys_state *css, int cpu)
struct mem_cgroup *memcg = mem_cgroup_from_css(css);
struct mem_cgroup *parent = parent_mem_cgroup(memcg);
struct memcg_vmstats_percpu *statc;
+ int *stats_updates_percpu;
long delta, delta_cpu, v;
int i, nid;
statc = per_cpu_ptr(memcg->vmstats_percpu, cpu);
+ stats_updates_percpu = per_cpu_ptr(memcg->stats_updates_percpu, cpu);
for (i = 0; i < MEMCG_NR_STAT; i++) {
/*
@@ -5826,10 +5828,10 @@ static void mem_cgroup_css_rstat_flush(struct
cgroup_subsys_state *css, int cpu)
}
}
}
- statc->stats_updates = 0;
+ *stats_updates_percpu = 0;
/* We are in a per-cpu loop here, only do the atomic write once */
- if (atomic64_read(&memcg->vmstats->stats_updates))
- atomic64_set(&memcg->vmstats->stats_updates, 0);
+ if (atomic64_read(&memcg->stats_updates))
+ atomic64_set(&memcg->stats_updates, 0);
}
#ifdef CONFIG_MMU
