Hi!
This is version 3 of this series. Version 2 can be found here:
https://lore.kernel.org/lkml/[email protected]
Wangyang and Arjan reported a bottleneck in the networking code related to
struct dst_entry::__refcnt. Performance tanks massively when concurrency on
a dst_entry increases.
This happens when there are a large amount of connections to or from the
same IP address. The memtier benchmark when run on the same host as
memcached amplifies this massively. But even over real network connections
this issue can be observed at an obviously smaller scale (due to the
network bandwith limitations in my setup, i.e. 1Gb). How to reproduce:
Run memcached with -t $N and memtier_benchmark with -t $M and --ratio=1:100
on the same machine. localhost connections amplify the problem.
Start with the defaults for $N and $M and increase them. Depending on
your machine this will tank at some point. But even in reasonably small
$N, $M scenarios the refcount operations and the resulting false sharing
fallout becomes visible in perf top. At some point it becomes the
dominating issue.
There are two factors which make this reference count a scalability issue:
1) False sharing
dst_entry:__refcnt is located at offset 64 of dst_entry, which puts
it into a seperate cacheline vs. the read mostly members located at
the beginning of the struct.
That prevents false sharing vs. the struct members in the first 64
bytes of the structure, but there is also
dst_entry::lwtstate
which is located after the reference count and in the same cache
line. This member is read after a reference count has been acquired.
The other problem is struct rtable, which embeds a struct dst_entry
at offset 0. struct dst_entry has a size of 112 bytes, which means
that the struct members of rtable which follow the dst member share
the same cache line as dst_entry::__refcnt. Especially
rtable::rt_genid
is also read by the contexts which have a reference count acquired
already.
When dst_entry:__refcnt is incremented or decremented via an atomic
operation these read accesses stall and contribute to the performance
problem.
2) atomic_inc_not_zero()
A reference on dst_entry:__refcnt is acquired via
atomic_inc_not_zero() and released via atomic_dec_return().
atomic_inc_not_zero() is implemted via a atomic_try_cmpxchg() loop,
which exposes O(N^2) behaviour under contention with N concurrent
operations. Contention scalability is degrading with even a small
amount of contenders and gets worse from there.
Lightweight instrumentation exposed an average of 8!! retry loops per
atomic_inc_not_zero() invocation in a inc()/dec() loop running
concurrently on 112 CPUs.
There is nothing which can be done to make atomic_inc_not_zero() more
scalable.
The following series addresses these issues:
1) Reorder and pad struct dst_entry to prevent the false sharing.
2) Implement and use a reference count implementation which avoids the
atomic_inc_not_zero() problem.
It is slightly less performant in the case of the final 0 -> -1
transition, but the deconstruction of these objects is a low
frequency event. get()/put() pairs are in the hotpath and that's
what this implementation optimizes for.
The algorithm of this reference count is only suitable for RCU
managed objects. Therefore it cannot replace the refcount_t
algorithm, which is also based on atomic_inc_not_zero(), due to a
subtle race condition related to the 0 -> -1 transition and the final
verdict to mark the reference count dead. See details in patch 2/3.
It might be just my lack of imagination which declares this to be
impossible and I'd be happy to be proven wrong.
As a bonus the new rcuref implementation provides underflow/overflow
detection and mitigation while being performance wise on par with
open coded atomic_inc_not_zero() / atomic_dec_return() pairs even in
the non-contended case.
The combination of these two changes results in performance gains in micro
benchmarks and also localhost and networked memtier benchmarks talking to
memcached. It's hard to quantify the benchmark results as they depend
heavily on the micro-architecture and the number of concurrent operations.
The overall gain of both changes for localhost memtier ranges from 1.2X to
3.2X and from +2% to %5% range for networked operations on a 1Gb connection.
A micro benchmark which enforces maximized concurrency shows a gain between
1.2X and 4.7X!!!
Obviously this is focussed on a particular problem and therefore needs to
be discussed in detail. It also requires wider testing outside of the cases
which this is focussed on.
Though the false sharing issue is obvious and should be addressed
independent of the more focussed reference count changes.
The series is also available from git:
git://git.kernel.org/pub/scm/linux/kernel/git/tglx/devel.git rcuref
Changes vs. V2:
- Rename __refcnt to __rcuref (Linus)
- Fix comments and changelogs (Mark, Qiuxu)
- Fixup kernel doc of generated atomic_add_negative() variants
I want to say thanks to Wangyang who analyzed the issue and provided the
initial fix for the false sharing problem. Further thanks go to Arjan
Peter, Marc, Will and Borislav for valuable input and providing test
results on machines which I do not have access to, and to Linus and
Eric, Qiuxu and Mark for helpful feedback.
Thanks,
tglx
I've stuffed the two atomic patches in tip/locking/rcuref for Jakub
which I then merged into tip/locking/core.
Jakub, you should be able to merge that topic branch (rc1 based) and
stuff the network bits on top.
If anything went sideways, please holler!
Hello:
This series was applied to netdev/net-next.git (main)
by Jakub Kicinski <[email protected]>:
On Thu, 23 Mar 2023 21:55:27 +0100 (CET) you wrote:
> Hi!
>
> This is version 3 of this series. Version 2 can be found here:
>
> https://lore.kernel.org/lkml/[email protected]
>
> Wangyang and Arjan reported a bottleneck in the networking code related to
> struct dst_entry::__refcnt. Performance tanks massively when concurrency on
> a dst_entry increases.
>
> [...]
Here is the summary with links:
- [V3,1/4,V2,1/4] net: dst: Prevent false sharing vs. dst_entry:: __refcnt
https://git.kernel.org/netdev/net-next/c/d288a162dd1c
- [V3,2/4] atomics: Provide atomic_add_negative() variants
https://git.kernel.org/netdev/net-next/c/e5ab9eff46b0
- [V3,3/4] atomics: Provide rcuref - scalable reference counting
https://git.kernel.org/netdev/net-next/c/ee1ee6db0779
- [V3,4/4] net: dst: Switch to rcuref_t reference counting
https://git.kernel.org/netdev/net-next/c/bc9d3a9f2afc
You are awesome, thank you!
--
Deet-doot-dot, I am a bot.
https://korg.docs.kernel.org/patchwork/pwbot.html
On Thu, Mar 23, 2023 at 09:55:27PM +0100, Thomas Gleixner wrote:
> Hi!
>
> This is version 3 of this series. Version 2 can be found here:
>
> https://lore.kernel.org/lkml/[email protected]
>
> Wangyang and Arjan reported a bottleneck in the networking code related to
> struct dst_entry::__refcnt. Performance tanks massively when concurrency on
> a dst_entry increases.
>
> This happens when there are a large amount of connections to or from the
> same IP address. The memtier benchmark when run on the same host as
> memcached amplifies this massively. But even over real network connections
> this issue can be observed at an obviously smaller scale (due to the
> network bandwith limitations in my setup, i.e. 1Gb). How to reproduce:
>
> Run memcached with -t $N and memtier_benchmark with -t $M and --ratio=1:100
> on the same machine. localhost connections amplify the problem.
>
> Start with the defaults for $N and $M and increase them. Depending on
> your machine this will tank at some point. But even in reasonably small
> $N, $M scenarios the refcount operations and the resulting false sharing
> fallout becomes visible in perf top. At some point it becomes the
> dominating issue.
>
> There are two factors which make this reference count a scalability issue:
>
> 1) False sharing
>
> dst_entry:__refcnt is located at offset 64 of dst_entry, which puts
> it into a seperate cacheline vs. the read mostly members located at
> the beginning of the struct.
>
> That prevents false sharing vs. the struct members in the first 64
> bytes of the structure, but there is also
>
> dst_entry::lwtstate
>
> which is located after the reference count and in the same cache
> line. This member is read after a reference count has been acquired.
>
> The other problem is struct rtable, which embeds a struct dst_entry
> at offset 0. struct dst_entry has a size of 112 bytes, which means
> that the struct members of rtable which follow the dst member share
> the same cache line as dst_entry::__refcnt. Especially
>
> rtable::rt_genid
>
> is also read by the contexts which have a reference count acquired
> already.
>
> When dst_entry:__refcnt is incremented or decremented via an atomic
> operation these read accesses stall and contribute to the performance
> problem.
>
> 2) atomic_inc_not_zero()
>
> A reference on dst_entry:__refcnt is acquired via
> atomic_inc_not_zero() and released via atomic_dec_return().
>
> atomic_inc_not_zero() is implemted via a atomic_try_cmpxchg() loop,
> which exposes O(N^2) behaviour under contention with N concurrent
> operations. Contention scalability is degrading with even a small
> amount of contenders and gets worse from there.
>
> Lightweight instrumentation exposed an average of 8!! retry loops per
> atomic_inc_not_zero() invocation in a inc()/dec() loop running
> concurrently on 112 CPUs.
Huh. 8 is pretty bad, 8! far worse, but 8!!? 3.4e168186???
(Sorry, couldn't resist...)
> There is nothing which can be done to make atomic_inc_not_zero() more
> scalable.
>
> The following series addresses these issues:
>
> 1) Reorder and pad struct dst_entry to prevent the false sharing.
>
> 2) Implement and use a reference count implementation which avoids the
> atomic_inc_not_zero() problem.
>
> It is slightly less performant in the case of the final 0 -> -1
> transition, but the deconstruction of these objects is a low
> frequency event. get()/put() pairs are in the hotpath and that's
> what this implementation optimizes for.
>
> The algorithm of this reference count is only suitable for RCU
> managed objects. Therefore it cannot replace the refcount_t
> algorithm, which is also based on atomic_inc_not_zero(), due to a
> subtle race condition related to the 0 -> -1 transition and the final
> verdict to mark the reference count dead. See details in patch 2/3.
>
> It might be just my lack of imagination which declares this to be
> impossible and I'd be happy to be proven wrong.
It is possible to make something like rcuref_get that does only a
READ_ONCE(), WRITE_ONCE() to storage local to the task, smp_mb(),
READ_ONCE() and compare in the common case. There would be something
like rcuref_put() that did an smp_store_release() to the same storage
local to the task.
Of course, there is always a catch, and here there are several:
1. Instead of just returning a pointer to a struct dst_entry,
sk_dst_get() would need to hand back both that pointer along
with a pointer to the aforementioned storage local to the task.
2. A generally useful implementation would require the counterpart
to rcuref_get() to sometimes allocate memory. Failure to allocate
memory would imply failure to gain a reference.
3. The structure would not need to be RCU-freed, but it still
would need to be freed specially. (All of the non-NULL
locations in all the storage local to the tasks needs to be
checked, but batching optimizations work well here.)
4. The smp_mb() compares well to the atomic_add_negative_relaxed(),
but getting rid of that smp_mb() either adds more RCU-like delays
on the free path on the one hand, or requires IPIs on the free
path on the other.
5. The aforementioned storage local to the task could instead be
local to the CPU, but at the expense of some put-side cache
misses and possible false sharing in the case where rcuref_get()
runs on one CPU and rcuref_put() runs on another. Of course,
the current rcuref_put() gets that already due to the use of
atomic_add_negative_release() on shared memory, so perhaps not
a big deal.
Not sure it is worth it, but you did ask!
If this does turn out to be an attractive option, please let me know.
Thanx, Paul
> As a bonus the new rcuref implementation provides underflow/overflow
> detection and mitigation while being performance wise on par with
> open coded atomic_inc_not_zero() / atomic_dec_return() pairs even in
> the non-contended case.
>
> The combination of these two changes results in performance gains in micro
> benchmarks and also localhost and networked memtier benchmarks talking to
> memcached. It's hard to quantify the benchmark results as they depend
> heavily on the micro-architecture and the number of concurrent operations.
>
> The overall gain of both changes for localhost memtier ranges from 1.2X to
> 3.2X and from +2% to %5% range for networked operations on a 1Gb connection.
>
> A micro benchmark which enforces maximized concurrency shows a gain between
> 1.2X and 4.7X!!!
>
> Obviously this is focussed on a particular problem and therefore needs to
> be discussed in detail. It also requires wider testing outside of the cases
> which this is focussed on.
>
> Though the false sharing issue is obvious and should be addressed
> independent of the more focussed reference count changes.
>
> The series is also available from git:
>
> git://git.kernel.org/pub/scm/linux/kernel/git/tglx/devel.git rcuref
>
> Changes vs. V2:
>
> - Rename __refcnt to __rcuref (Linus)
>
> - Fix comments and changelogs (Mark, Qiuxu)
>
> - Fixup kernel doc of generated atomic_add_negative() variants
>
> I want to say thanks to Wangyang who analyzed the issue and provided the
> initial fix for the false sharing problem. Further thanks go to Arjan
> Peter, Marc, Will and Borislav for valuable input and providing test
> results on machines which I do not have access to, and to Linus and
> Eric, Qiuxu and Mark for helpful feedback.
>
> Thanks,
>
> tglx
>
On Thu, Mar 23, 2023 at 09:55:27PM +0100, Thomas Gleixner wrote:
> Hi!
>
> This is version 3 of this series. Version 2 can be found here:
>
> https://lore.kernel.org/lkml/[email protected]
Hi,
I want to raise your attention to this bug report from Intel.
https://lore.kernel.org/all/[email protected]/
We (Nvidia) are experiencing similar failures in our regressions too.
Revert of last two patches [1] from this series removed the panics, but
didn't add confidence due to another (???) netdev failure:
[ +10.080020] unregister_netdevice: waiting for eth3 to become free. Usage count = 2
Thanks
[1]
bc9d3a9f2afc net: dst: Switch to rcuref_t reference counting
d288a162dd1c net: dst: Prevent false sharing vs. dst_entry:: __refcnt