We've noticed a quite sensible performance overhead on some hosts
with significant network traffic when socket memory accounting
is enabled.
Perf top shows that socket memory uncharging path is hot:
2.13% [kernel] [k] page_counter_cancel
1.14% [kernel] [k] __sk_mem_reduce_allocated
1.14% [kernel] [k] _raw_spin_lock
0.87% [kernel] [k] _raw_spin_lock_irqsave
0.84% [kernel] [k] tcp_ack
0.84% [kernel] [k] ixgbe_poll
0.83% < workload >
0.82% [kernel] [k] enqueue_entity
0.68% [kernel] [k] __fget
0.68% [kernel] [k] tcp_delack_timer_handler
0.67% [kernel] [k] __schedule
0.60% < workload >
0.59% [kernel] [k] __inet6_lookup_established
0.55% [kernel] [k] __switch_to
0.55% [kernel] [k] menu_select
0.54% libc-2.20.so [.] __memcpy_avx_unaligned
To address this issue, the existing per-cpu stock infrastructure
can be used.
refill_stock() can be called from mem_cgroup_uncharge_skmem()
to move charge to a per-cpu stock instead of calling atomic
page_counter_uncharge().
To prevent the uncontrolled growth of per-cpu stocks,
refill_stock() will explicitly drain the cached charge,
if the cached value exceeds CHARGE_BATCH.
This allows significantly optimize the load:
1.21% [kernel] [k] _raw_spin_lock
1.01% [kernel] [k] ixgbe_poll
0.92% [kernel] [k] _raw_spin_lock_irqsave
0.90% [kernel] [k] enqueue_entity
0.86% [kernel] [k] tcp_ack
0.85% < workload >
0.74% perf-11120.map [.] 0x000000000061bf24
0.73% [kernel] [k] __schedule
0.67% [kernel] [k] __fget
0.63% [kernel] [k] __inet6_lookup_established
0.62% [kernel] [k] menu_select
0.59% < workload >
0.59% [kernel] [k] __switch_to
0.57% libc-2.20.so [.] __memcpy_avx_unaligned
Signed-off-by: Roman Gushchin <[email protected]>
Cc: Johannes Weiner <[email protected]>
Cc: Michal Hocko <[email protected]>
Cc: Vladimir Davydov <[email protected]>
Cc: [email protected]
Cc: [email protected]
Cc: [email protected]
Cc: [email protected]
---
mm/memcontrol.c | 6 ++++--
1 file changed, 4 insertions(+), 2 deletions(-)
diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index b9cf3cf4a3d0..a69d23082abf 100644
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -1792,6 +1792,9 @@ static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
}
stock->nr_pages += nr_pages;
+ if (stock->nr_pages > CHARGE_BATCH)
+ drain_stock(stock);
+
local_irq_restore(flags);
}
@@ -5886,8 +5889,7 @@ void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
this_cpu_sub(memcg->stat->count[MEMCG_SOCK], nr_pages);
- page_counter_uncharge(&memcg->memory, nr_pages);
- css_put_many(&memcg->css, nr_pages);
+ refill_stock(memcg, nr_pages);
}
static int __init cgroup_memory(char *s)
--
2.13.5
On Tue, Aug 29, 2017 at 11:01:50AM +0100, Roman Gushchin wrote:
> We've noticed a quite sensible performance overhead on some hosts
> with significant network traffic when socket memory accounting
> is enabled.
>
> Perf top shows that socket memory uncharging path is hot:
> 2.13% [kernel] [k] page_counter_cancel
> 1.14% [kernel] [k] __sk_mem_reduce_allocated
> 1.14% [kernel] [k] _raw_spin_lock
> 0.87% [kernel] [k] _raw_spin_lock_irqsave
> 0.84% [kernel] [k] tcp_ack
> 0.84% [kernel] [k] ixgbe_poll
> 0.83% < workload >
> 0.82% [kernel] [k] enqueue_entity
> 0.68% [kernel] [k] __fget
> 0.68% [kernel] [k] tcp_delack_timer_handler
> 0.67% [kernel] [k] __schedule
> 0.60% < workload >
> 0.59% [kernel] [k] __inet6_lookup_established
> 0.55% [kernel] [k] __switch_to
> 0.55% [kernel] [k] menu_select
> 0.54% libc-2.20.so [.] __memcpy_avx_unaligned
>
> To address this issue, the existing per-cpu stock infrastructure
> can be used.
>
> refill_stock() can be called from mem_cgroup_uncharge_skmem()
> to move charge to a per-cpu stock instead of calling atomic
> page_counter_uncharge().
>
> To prevent the uncontrolled growth of per-cpu stocks,
> refill_stock() will explicitly drain the cached charge,
> if the cached value exceeds CHARGE_BATCH.
>
> This allows significantly optimize the load:
> 1.21% [kernel] [k] _raw_spin_lock
> 1.01% [kernel] [k] ixgbe_poll
> 0.92% [kernel] [k] _raw_spin_lock_irqsave
> 0.90% [kernel] [k] enqueue_entity
> 0.86% [kernel] [k] tcp_ack
> 0.85% < workload >
> 0.74% perf-11120.map [.] 0x000000000061bf24
> 0.73% [kernel] [k] __schedule
> 0.67% [kernel] [k] __fget
> 0.63% [kernel] [k] __inet6_lookup_established
> 0.62% [kernel] [k] menu_select
> 0.59% < workload >
> 0.59% [kernel] [k] __switch_to
> 0.57% libc-2.20.so [.] __memcpy_avx_unaligned
>
> Signed-off-by: Roman Gushchin <[email protected]>
> Cc: Johannes Weiner <[email protected]>
> Cc: Michal Hocko <[email protected]>
> Cc: Vladimir Davydov <[email protected]>
> Cc: [email protected]
> Cc: [email protected]
> Cc: [email protected]
> Cc: [email protected]
Acked-by: Johannes Weiner <[email protected]>
Neat!
As far as other types of pages go: page cache and anon are already
batched pretty well, but I think kmem might benefit from this
too. Have you considered using the stock in memcg_kmem_uncharge()?
On Tue, Aug 29, 2017 at 03:26:21PM -0400, Johannes Weiner wrote:
> On Tue, Aug 29, 2017 at 11:01:50AM +0100, Roman Gushchin wrote:
> > We've noticed a quite sensible performance overhead on some hosts
> > with significant network traffic when socket memory accounting
> > is enabled.
> >
> > Perf top shows that socket memory uncharging path is hot:
> > 2.13% [kernel] [k] page_counter_cancel
> > 1.14% [kernel] [k] __sk_mem_reduce_allocated
> > 1.14% [kernel] [k] _raw_spin_lock
> > 0.87% [kernel] [k] _raw_spin_lock_irqsave
> > 0.84% [kernel] [k] tcp_ack
> > 0.84% [kernel] [k] ixgbe_poll
> > 0.83% < workload >
> > 0.82% [kernel] [k] enqueue_entity
> > 0.68% [kernel] [k] __fget
> > 0.68% [kernel] [k] tcp_delack_timer_handler
> > 0.67% [kernel] [k] __schedule
> > 0.60% < workload >
> > 0.59% [kernel] [k] __inet6_lookup_established
> > 0.55% [kernel] [k] __switch_to
> > 0.55% [kernel] [k] menu_select
> > 0.54% libc-2.20.so [.] __memcpy_avx_unaligned
> >
> > To address this issue, the existing per-cpu stock infrastructure
> > can be used.
> >
> > refill_stock() can be called from mem_cgroup_uncharge_skmem()
> > to move charge to a per-cpu stock instead of calling atomic
> > page_counter_uncharge().
> >
> > To prevent the uncontrolled growth of per-cpu stocks,
> > refill_stock() will explicitly drain the cached charge,
> > if the cached value exceeds CHARGE_BATCH.
> >
> > This allows significantly optimize the load:
> > 1.21% [kernel] [k] _raw_spin_lock
> > 1.01% [kernel] [k] ixgbe_poll
> > 0.92% [kernel] [k] _raw_spin_lock_irqsave
> > 0.90% [kernel] [k] enqueue_entity
> > 0.86% [kernel] [k] tcp_ack
> > 0.85% < workload >
> > 0.74% perf-11120.map [.] 0x000000000061bf24
> > 0.73% [kernel] [k] __schedule
> > 0.67% [kernel] [k] __fget
> > 0.63% [kernel] [k] __inet6_lookup_established
> > 0.62% [kernel] [k] menu_select
> > 0.59% < workload >
> > 0.59% [kernel] [k] __switch_to
> > 0.57% libc-2.20.so [.] __memcpy_avx_unaligned
> >
> > Signed-off-by: Roman Gushchin <[email protected]>
> > Cc: Johannes Weiner <[email protected]>
> > Cc: Michal Hocko <[email protected]>
> > Cc: Vladimir Davydov <[email protected]>
> > Cc: [email protected]
> > Cc: [email protected]
> > Cc: [email protected]
> > Cc: [email protected]
>
> Acked-by: Johannes Weiner <[email protected]>
>
> Neat!
>
> As far as other types of pages go: page cache and anon are already
> batched pretty well, but I think kmem might benefit from this
> too. Have you considered using the stock in memcg_kmem_uncharge()?
Good idea!
I'll try to find an appropriate testcase and check if it really
brings any benefits. If so, I'll master a patch.
Thanks!
On Tue, Aug 29, 2017 at 03:26:21PM -0400, Johannes Weiner wrote:
> On Tue, Aug 29, 2017 at 11:01:50AM +0100, Roman Gushchin wrote:
> > We've noticed a quite sensible performance overhead on some hosts
> > with significant network traffic when socket memory accounting
> > is enabled.
> >
> > Perf top shows that socket memory uncharging path is hot:
> > 2.13% [kernel] [k] page_counter_cancel
> > 1.14% [kernel] [k] __sk_mem_reduce_allocated
> > 1.14% [kernel] [k] _raw_spin_lock
> > 0.87% [kernel] [k] _raw_spin_lock_irqsave
> > 0.84% [kernel] [k] tcp_ack
> > 0.84% [kernel] [k] ixgbe_poll
> > 0.83% < workload >
> > 0.82% [kernel] [k] enqueue_entity
> > 0.68% [kernel] [k] __fget
> > 0.68% [kernel] [k] tcp_delack_timer_handler
> > 0.67% [kernel] [k] __schedule
> > 0.60% < workload >
> > 0.59% [kernel] [k] __inet6_lookup_established
> > 0.55% [kernel] [k] __switch_to
> > 0.55% [kernel] [k] menu_select
> > 0.54% libc-2.20.so [.] __memcpy_avx_unaligned
> >
> > To address this issue, the existing per-cpu stock infrastructure
> > can be used.
> >
> > refill_stock() can be called from mem_cgroup_uncharge_skmem()
> > to move charge to a per-cpu stock instead of calling atomic
> > page_counter_uncharge().
> >
> > To prevent the uncontrolled growth of per-cpu stocks,
> > refill_stock() will explicitly drain the cached charge,
> > if the cached value exceeds CHARGE_BATCH.
> >
> > This allows significantly optimize the load:
> > 1.21% [kernel] [k] _raw_spin_lock
> > 1.01% [kernel] [k] ixgbe_poll
> > 0.92% [kernel] [k] _raw_spin_lock_irqsave
> > 0.90% [kernel] [k] enqueue_entity
> > 0.86% [kernel] [k] tcp_ack
> > 0.85% < workload >
> > 0.74% perf-11120.map [.] 0x000000000061bf24
> > 0.73% [kernel] [k] __schedule
> > 0.67% [kernel] [k] __fget
> > 0.63% [kernel] [k] __inet6_lookup_established
> > 0.62% [kernel] [k] menu_select
> > 0.59% < workload >
> > 0.59% [kernel] [k] __switch_to
> > 0.57% libc-2.20.so [.] __memcpy_avx_unaligned
> >
> > Signed-off-by: Roman Gushchin <[email protected]>
> > Cc: Johannes Weiner <[email protected]>
> > Cc: Michal Hocko <[email protected]>
> > Cc: Vladimir Davydov <[email protected]>
> > Cc: [email protected]
> > Cc: [email protected]
> > Cc: [email protected]
> > Cc: [email protected]
>
> Acked-by: Johannes Weiner <[email protected]>
>
> Neat!
Hi, Andrew!
Can you, please, pull this patch?
Thank you!
Roman
On Tue 29-08-17 11:01:50, Roman Gushchin wrote:
[...]
> diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> index b9cf3cf4a3d0..a69d23082abf 100644
> --- a/mm/memcontrol.c
> +++ b/mm/memcontrol.c
> @@ -1792,6 +1792,9 @@ static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
> }
> stock->nr_pages += nr_pages;
>
> + if (stock->nr_pages > CHARGE_BATCH)
> + drain_stock(stock);
> +
> local_irq_restore(flags);
> }
Why do we need this? In other words, why cannot we rely on draining we
already do?
>
> @@ -5886,8 +5889,7 @@ void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
>
> this_cpu_sub(memcg->stat->count[MEMCG_SOCK], nr_pages);
>
> - page_counter_uncharge(&memcg->memory, nr_pages);
> - css_put_many(&memcg->css, nr_pages);
> + refill_stock(memcg, nr_pages);
> }
This makes sense to me.
--
Michal Hocko
SUSE Labs
On Wed, Aug 30, 2017 at 02:36:55PM +0200, Michal Hocko wrote:
> On Tue 29-08-17 11:01:50, Roman Gushchin wrote:
> [...]
> > diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> > index b9cf3cf4a3d0..a69d23082abf 100644
> > --- a/mm/memcontrol.c
> > +++ b/mm/memcontrol.c
> > @@ -1792,6 +1792,9 @@ static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
> > }
> > stock->nr_pages += nr_pages;
> >
> > + if (stock->nr_pages > CHARGE_BATCH)
> > + drain_stock(stock);
> > +
> > local_irq_restore(flags);
> > }
>
> Why do we need this? In other words, why cannot we rely on draining we
> already do?
The existing draining depends on memory pressure, so to keep
the accounting (which we expose to a user) reasonable accurate
even without memory pressure, we need to limit the size of per-cpu stocks.
Thanks!
On Wed 30-08-17 13:44:59, Roman Gushchin wrote:
> On Wed, Aug 30, 2017 at 02:36:55PM +0200, Michal Hocko wrote:
> > On Tue 29-08-17 11:01:50, Roman Gushchin wrote:
> > [...]
> > > diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> > > index b9cf3cf4a3d0..a69d23082abf 100644
> > > --- a/mm/memcontrol.c
> > > +++ b/mm/memcontrol.c
> > > @@ -1792,6 +1792,9 @@ static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
> > > }
> > > stock->nr_pages += nr_pages;
> > >
> > > + if (stock->nr_pages > CHARGE_BATCH)
> > > + drain_stock(stock);
> > > +
> > > local_irq_restore(flags);
> > > }
> >
> > Why do we need this? In other words, why cannot we rely on draining we
> > already do?
>
> The existing draining depends on memory pressure, so to keep
> the accounting (which we expose to a user) reasonable accurate
> even without memory pressure, we need to limit the size of per-cpu stocks.
Why don't we need this for regular page charges? Or maybe we do but that
sounds like a seprate and an unrealted fix to me.
--
Michal Hocko
SUSE Labs
On Wed, Aug 30, 2017 at 02:55:43PM +0200, Michal Hocko wrote:
> On Wed 30-08-17 13:44:59, Roman Gushchin wrote:
> > On Wed, Aug 30, 2017 at 02:36:55PM +0200, Michal Hocko wrote:
> > > On Tue 29-08-17 11:01:50, Roman Gushchin wrote:
> > > [...]
> > > > diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> > > > index b9cf3cf4a3d0..a69d23082abf 100644
> > > > --- a/mm/memcontrol.c
> > > > +++ b/mm/memcontrol.c
> > > > @@ -1792,6 +1792,9 @@ static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
> > > > }
> > > > stock->nr_pages += nr_pages;
> > > >
> > > > + if (stock->nr_pages > CHARGE_BATCH)
> > > > + drain_stock(stock);
> > > > +
> > > > local_irq_restore(flags);
> > > > }
> > >
> > > Why do we need this? In other words, why cannot we rely on draining we
> > > already do?
> >
> > The existing draining depends on memory pressure, so to keep
> > the accounting (which we expose to a user) reasonable accurate
> > even without memory pressure, we need to limit the size of per-cpu stocks.
>
> Why don't we need this for regular page charges? Or maybe we do but that
> sounds like a seprate and an unrealted fix to me.
Because we never refill more than CHARGE_BATCH.
On Wed 30-08-17 13:57:29, Roman Gushchin wrote:
> On Wed, Aug 30, 2017 at 02:55:43PM +0200, Michal Hocko wrote:
> > On Wed 30-08-17 13:44:59, Roman Gushchin wrote:
> > > On Wed, Aug 30, 2017 at 02:36:55PM +0200, Michal Hocko wrote:
> > > > On Tue 29-08-17 11:01:50, Roman Gushchin wrote:
> > > > [...]
> > > > > diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> > > > > index b9cf3cf4a3d0..a69d23082abf 100644
> > > > > --- a/mm/memcontrol.c
> > > > > +++ b/mm/memcontrol.c
> > > > > @@ -1792,6 +1792,9 @@ static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
> > > > > }
> > > > > stock->nr_pages += nr_pages;
> > > > >
> > > > > + if (stock->nr_pages > CHARGE_BATCH)
> > > > > + drain_stock(stock);
> > > > > +
> > > > > local_irq_restore(flags);
> > > > > }
> > > >
> > > > Why do we need this? In other words, why cannot we rely on draining we
> > > > already do?
> > >
> > > The existing draining depends on memory pressure, so to keep
> > > the accounting (which we expose to a user) reasonable accurate
> > > even without memory pressure, we need to limit the size of per-cpu stocks.
> >
> > Why don't we need this for regular page charges? Or maybe we do but that
> > sounds like a seprate and an unrealted fix to me.
>
> Because we never refill more than CHARGE_BATCH.
You are right that a single process will not but try_charge is a preemptible
context and so multiple processes might pass consume_stock and then
charge a N*CHARGE_BATCH. But I agree that this is quite unlikely so a
separate patch is probably not worth it.
--
Michal Hocko
SUSE Labs
On Tue 29-08-17 11:01:50, Roman Gushchin wrote:
> We've noticed a quite sensible performance overhead on some hosts
> with significant network traffic when socket memory accounting
> is enabled.
>
> Perf top shows that socket memory uncharging path is hot:
> 2.13% [kernel] [k] page_counter_cancel
> 1.14% [kernel] [k] __sk_mem_reduce_allocated
> 1.14% [kernel] [k] _raw_spin_lock
> 0.87% [kernel] [k] _raw_spin_lock_irqsave
> 0.84% [kernel] [k] tcp_ack
> 0.84% [kernel] [k] ixgbe_poll
> 0.83% < workload >
> 0.82% [kernel] [k] enqueue_entity
> 0.68% [kernel] [k] __fget
> 0.68% [kernel] [k] tcp_delack_timer_handler
> 0.67% [kernel] [k] __schedule
> 0.60% < workload >
> 0.59% [kernel] [k] __inet6_lookup_established
> 0.55% [kernel] [k] __switch_to
> 0.55% [kernel] [k] menu_select
> 0.54% libc-2.20.so [.] __memcpy_avx_unaligned
>
> To address this issue, the existing per-cpu stock infrastructure
> can be used.
>
> refill_stock() can be called from mem_cgroup_uncharge_skmem()
> to move charge to a per-cpu stock instead of calling atomic
> page_counter_uncharge().
>
> To prevent the uncontrolled growth of per-cpu stocks,
> refill_stock() will explicitly drain the cached charge,
> if the cached value exceeds CHARGE_BATCH.
>
> This allows significantly optimize the load:
> 1.21% [kernel] [k] _raw_spin_lock
> 1.01% [kernel] [k] ixgbe_poll
> 0.92% [kernel] [k] _raw_spin_lock_irqsave
> 0.90% [kernel] [k] enqueue_entity
> 0.86% [kernel] [k] tcp_ack
> 0.85% < workload >
> 0.74% perf-11120.map [.] 0x000000000061bf24
> 0.73% [kernel] [k] __schedule
> 0.67% [kernel] [k] __fget
> 0.63% [kernel] [k] __inet6_lookup_established
> 0.62% [kernel] [k] menu_select
> 0.59% < workload >
> 0.59% [kernel] [k] __switch_to
> 0.57% libc-2.20.so [.] __memcpy_avx_unaligned
>
> Signed-off-by: Roman Gushchin <[email protected]>
> Cc: Johannes Weiner <[email protected]>
> Cc: Michal Hocko <[email protected]>
> Cc: Vladimir Davydov <[email protected]>
> Cc: [email protected]
> Cc: [email protected]
> Cc: [email protected]
> Cc: [email protected]
Acked-by: Michal Hocko <[email protected]>
> ---
> mm/memcontrol.c | 6 ++++--
> 1 file changed, 4 insertions(+), 2 deletions(-)
>
> diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> index b9cf3cf4a3d0..a69d23082abf 100644
> --- a/mm/memcontrol.c
> +++ b/mm/memcontrol.c
> @@ -1792,6 +1792,9 @@ static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
> }
> stock->nr_pages += nr_pages;
>
> + if (stock->nr_pages > CHARGE_BATCH)
> + drain_stock(stock);
> +
> local_irq_restore(flags);
> }
>
> @@ -5886,8 +5889,7 @@ void mem_cgroup_uncharge_skmem(struct mem_cgroup *memcg, unsigned int nr_pages)
>
> this_cpu_sub(memcg->stat->count[MEMCG_SOCK], nr_pages);
>
> - page_counter_uncharge(&memcg->memory, nr_pages);
> - css_put_many(&memcg->css, nr_pages);
> + refill_stock(memcg, nr_pages);
> }
>
> static int __init cgroup_memory(char *s)
> --
> 2.13.5
--
Michal Hocko
SUSE Labs
>> As far as other types of pages go: page cache and anon are already
>> batched pretty well, but I think kmem might benefit from this
>> too. Have you considered using the stock in memcg_kmem_uncharge()?
>
> Good idea!
> I'll try to find an appropriate testcase and check if it really
> brings any benefits. If so, I'll master a patch.
>
Hi Roman, did you get the chance to try this on memcg_kmem_uncharge()?
On Thu, Sep 07, 2017 at 11:44:12AM -0700, Shakeel Butt wrote:
> >> As far as other types of pages go: page cache and anon are already
> >> batched pretty well, but I think kmem might benefit from this
> >> too. Have you considered using the stock in memcg_kmem_uncharge()?
> >
> > Good idea!
> > I'll try to find an appropriate testcase and check if it really
> > brings any benefits. If so, I'll master a patch.
> >
>
> Hi Roman, did you get the chance to try this on memcg_kmem_uncharge()?
Hi Shakeel!
Not yet, I'll try to it asap.
Do you have an example when it's costly?
Thanks!
On Thu, Sep 7, 2017 at 11:47 AM, Roman Gushchin <[email protected]> wrote:
> On Thu, Sep 07, 2017 at 11:44:12AM -0700, Shakeel Butt wrote:
>> >> As far as other types of pages go: page cache and anon are already
>> >> batched pretty well, but I think kmem might benefit from this
>> >> too. Have you considered using the stock in memcg_kmem_uncharge()?
>> >
>> > Good idea!
>> > I'll try to find an appropriate testcase and check if it really
>> > brings any benefits. If so, I'll master a patch.
>> >
>>
>> Hi Roman, did you get the chance to try this on memcg_kmem_uncharge()?
>
> Hi Shakeel!
>
> Not yet, I'll try to it asap.
> Do you have an example when it's costly?
>
Nah, I was also curious of such examples.