Skip to contentSkip to conversation

Please backport de53fd7aedb1 : sched/fair: Fix low cpu usage with high throttling by removing expiration of cpu-local slices

PATCH In review Dave Chiluk 5 participants 20 messages Sep 25, 2019 – Nov 11, 2019

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

Commit de53fd7aedb1 ("sched/fair: Fix low cpu usage with high throttling by removing expiration of cpu-local slices") fixes a major performance issue for containerized clouds such as Kubernetes.

Commit de53fd7aedb1 Fixes commit : 512ac999d275 ("sched/fair: Fix bandwidth timer clock drift condition").

This should be applied to all stable kernels that applied commit 512ac999d275, and should probably be applied to all others as well.

The issues introduced by these pathes can be read about on the Kubernetes github.
https://github.com/kubernetes/kubernetes/issues/67577

It may also be prudent to also apply the not yet accepted patch that fixes some introduced compiler warnings discussed here.
https://lkml.org/lkml/2019/9/18/925

Thank you,
Dave Chiluk

#1 lore raw
Quoted 9 lines · Dave Chiluk

Commit de53fd7aedb1 ("sched/fair: Fix low cpu usage with high throttling by removing expiration of cpu-local slices") fixes a major performance issue for containerized clouds such as Kubernetes.

Commit de53fd7aedb1 Fixes commit : 512ac999d275 ("sched/fair: Fix bandwidth timer clock drift condition").

This should be applied to all stable kernels that applied commit 512ac999d275, and should probably be applied to all others as well.

As this commit isn't in a released kernel just yet, we should wait to see what happens when it hits people's machines, right?

Also, always cc: all of the people involved in the patch you are asking for, so as to get their opinion. For some reason this patch did not have a cc: stable tag on it, was that because the developers did not think it was relevant for stable kernels?

Quoted 7 lines

The issues introduced by these pathes can be read about on the Kubernetes github.
https://github.com/kubernetes/kubernetes/issues/67577

It may also be prudent to also apply the not yet accepted patch that fixes some introduced compiler warnings discussed here.
https://lkml.org/lkml/2019/9/18/925

So we should wait for this to hit Linus's tree too, right?

thanks,

greg k-h

#2 lore raw
Quoted 6 lines · Greg KH
Quoted 4 lines · Dave Chiluk

Commit de53fd7aedb1 ("sched/fair: Fix low cpu usage with high throttling by removing expiration of cpu-local slices") fixes a major performance issue for containerized clouds such as Kubernetes.

Quoted 8 lines
Quoted 5 lines

Commit de53fd7aedb1 Fixes commit : 512ac999d275 ("sched/fair: Fix bandwidth timer clock drift condition").

This should be applied to all stable kernels that applied commit 512ac999d275, and should probably be applied to all others as well.

As this commit isn't in a released kernel just yet, we should wait to see what happens when it hits people's machines, right?

I think waiting till 5.4 is released would be irresponsible on this one. 512ac999 was recently pushed back into many of the distro kernels via the stable streams. Additionally every major container running cloud provider (public and private), is likely hitting this whether they've noticed or not. Before 512ac999 we would hit the issue that resolved once every 10000 application deployments or so *(I admit some providers appear to have been hit by that more often). However, the fix implemented by 512ac999 is guaranteed to affect 100% of cgroup cpu bandwidth constrained applications.

Our java applications are particularly hard hit as java tends to be
very thread happy. Doing some napkin math, given the roughly 9000
applications we have running in our clouds we've had to over-allocate each one's CPU quota by roughly .5 cpu to account for this behavior change (worst case scenario is actually .01 cpu * cores in machine, but not all of our applications are affected equally).
9000 applications * .5 CPU = 4500 cores
4500 Cores / 88 cores per node = 51 additional machines required to satisfy the inflated quota requirements.
Given each 88 core machine costs roughly $30k that equates to $1.5M that this issue is costing us.
We have been able to alleviate this somewhat of this by turning on CPU overcommit on in the container scheduler. (Deploying more applications per host than the CPU would generally allow for). This however leads to occasional overloaded machines, and regular high response times/wide response time distributions for applications.

If you have a java or golang application running on a containerized cloud anywhere chances are you are either paying extra to get the CPU you need or your application is being throttled before using the quota you've paid for. I'm sure other languages could be affected, but these are the two we've seen that by default generate enough threads to regularly hit this issue.

All of these problems are leading the kubernetes community to change the defaulrs away from the cfs bandwidth scheduling mechanisms and towards other mechanisms or hackish workarounds. *(adjusting periods, turning off cfs bandwidth, moving to only soft limits).
https://github.com/kubernetes/kubernetes/issues/67577
https://github.com/kubernetes/kubernetes/issues/70585
https://github.com/kubernetes/kubernetes/issues/51135

I appreciate your reluctance to accept this patch, but I strongly believe pulling this sooner than later is the right thing to do.

Quoted 2 lines

Also, always cc: all of the people involved in the patch you are asking for, so as to get their opinion.

Fixed. I'll submit a change to stable-kernel-rules.rst upon my return from vacation.

Quoted 3 lines

For some reason this patch did not
have a cc: stable tag on it, was that because the developers did not think it was relevant for stable kernels?

This was my fault, I was unaware I could simply cc: stable to have it auto-submitted. I promise I'll be better in the future.

Quoted 5 lines
Quoted 3 lines

It may also be prudent to also apply the not yet accepted patch that fixes some introduced compiler warnings discussed here.
https://lkml.org/lkml/2019/9/18/925

So we should wait for this to hit Linus's tree too, right?

This patch contributes nothing materially to de53fd7aedb1, and only eliminates compiler warnings for unused variables *(which likely get optimized out anyway). I'm not sure what the stable policy is with introduced compiler warnings, so I included the patch submission in this discussion for completeness.

Thank you,
Dave Chiluk

#3 lore raw
Quoted 58 lines · Dave Chiluk
Quoted 6 lines · Greg KH
Quoted 4 lines · Dave Chiluk

Commit de53fd7aedb1 ("sched/fair: Fix low cpu usage with high throttling by removing expiration of cpu-local slices") fixes a major performance issue for containerized clouds such as Kubernetes.

Quoted 8 lines
Quoted 5 lines

Commit de53fd7aedb1 Fixes commit : 512ac999d275 ("sched/fair: Fix bandwidth timer clock drift condition").

This should be applied to all stable kernels that applied commit 512ac999d275, and should probably be applied to all others as well.

As this commit isn't in a released kernel just yet, we should wait to see what happens when it hits people's machines, right?

I think waiting till 5.4 is released would be irresponsible on this one. 512ac999 was recently pushed back into many of the distro kernels via the stable streams. Additionally every major container running cloud provider (public and private), is likely hitting this whether they've noticed or not. Before 512ac999 we would hit the issue that resolved once every 10000 application deployments or so *(I admit some providers appear to have been hit by that more often). However, the fix implemented by 512ac999 is guaranteed to affect 100% of cgroup cpu bandwidth constrained applications.

Our java applications are particularly hard hit as java tends to be
very thread happy. Doing some napkin math, given the roughly 9000
applications we have running in our clouds we've had to over-allocate each one's CPU quota by roughly .5 cpu to account for this behavior change (worst case scenario is actually .01 cpu * cores in machine, but not all of our applications are affected equally).
9000 applications * .5 CPU = 4500 cores
4500 Cores / 88 cores per node = 51 additional machines required to satisfy the inflated quota requirements.
Given each 88 core machine costs roughly $30k that equates to $1.5M that this issue is costing us.
We have been able to alleviate this somewhat of this by turning on CPU overcommit on in the container scheduler. (Deploying more applications per host than the CPU would generally allow for). This however leads to occasional overloaded machines, and regular high response times/wide response time distributions for applications.

If you have a java or golang application running on a containerized cloud anywhere chances are you are either paying extra to get the CPU you need or your application is being throttled before using the quota you've paid for. I'm sure other languages could be affected, but these are the two we've seen that by default generate enough threads to regularly hit this issue.

All of these problems are leading the kubernetes community to change the defaulrs away from the cfs bandwidth scheduling mechanisms and towards other mechanisms or hackish workarounds. *(adjusting periods, turning off cfs bandwidth, moving to only soft limits).
https://github.com/kubernetes/kubernetes/issues/67577
https://github.com/kubernetes/kubernetes/issues/70585
https://github.com/kubernetes/kubernetes/issues/51135

Quoted 2 lines

I appreciate your reluctance to accept this patch, but I strongly believe pulling this sooner than later is the right thing to do.

I'll defer to the scheduler developers/maintainers here as to when to take this patch, it's their decision.

Quoted 4 lines
Quoted 2 lines

Also, always cc: all of the people involved in the patch you are asking for, so as to get their opinion.

Fixed. I'll submit a change to stable-kernel-rules.rst upon my return from vacation.

Great!

Quoted 6 lines
Quoted 3 lines

For some reason this patch did not
have a cc: stable tag on it, was that because the developers did not think it was relevant for stable kernels?

This was my fault, I was unaware I could simply cc: stable to have it auto-submitted. I promise I'll be better in the future.

Even if this was "auto-submitted", it would not show up in a kernel until it hit a release of Linus's tree (i.e. 5.4-rc1.) So we would still be having this discussion :)

Quoted 11 lines
Quoted 5 lines
Quoted 3 lines

It may also be prudent to also apply the not yet accepted patch that fixes some introduced compiler warnings discussed here.
https://lkml.org/lkml/2019/9/18/925

So we should wait for this to hit Linus's tree too, right?

This patch contributes nothing materially to de53fd7aedb1, and only eliminates compiler warnings for unused variables *(which likely get optimized out anyway). I'm not sure what the stable policy is with introduced compiler warnings, so I included the patch submission in this discussion for completeness.

Adding build warnings is a huge no-no as it has the ability to hide real problems with backports. Right now the stable trees build for me with no, or only 1, warnings, depending on the branch. That is something that I want to see continue to happen for obvious reasons. So including a patch that we know adds warnings is not good.

It's also a huge hint that it didn't get much, if any, testing in linux-next as the warning would have shown up there, and hopefully fixed soon, as part of the original merge. But I digress...

So, I'll be glad to merge this "now" if I get an ACK from the others. Otherwise I'll have to wait until Monday.

thanks,

greg k-h

maintainer
#4 lore raw
In reply to Dave Chiluk
Quoted 10 lines · Dave Chiluk
Quoted 9 lines · Greg KH
Quoted 7 lines · Dave Chiluk

Commit de53fd7aedb1 ("sched/fair: Fix low cpu usage with high throttling by removing expiration of cpu-local slices") fixes a major performance issue for containerized clouds such as Kubernetes.

Commit de53fd7aedb1 Fixes commit : 512ac999d275 ("sched/fair: Fix bandwidth timer clock drift condition").

Quoted 8 lines
Quoted 5 lines
Quoted 2 lines

This should be applied to all stable kernels that applied commit 512ac999d275, and should probably be applied to all others as well.

As this commit isn't in a released kernel just yet, we should wait to see what happens when it hits people's machines, right?

I think waiting till 5.4 is released would be irresponsible on this one. 512ac999 was recently pushed back into many of the distro

The "released kernel" statement means 5.4-rc1 rather than 5.4, which is just a few days away.

--
Thanks,
Sasha

#5 lore raw

@Greg KH, Qian Cai's compiler warning fix has now been integrated into Linus' tree as commit: 763a9ec06c409

Both de53fd7aedb1 and 763a9ec06c40 are now apart of v5.4-rc1. Can you please queue up these fixes for backport to all stable kernels.

Thank you,
Dave Chiluk

Quoted 27 lines · Sasha Levin
Quoted 10 lines · Dave Chiluk
Quoted 9 lines · Greg KH
Quoted 7 lines · Dave Chiluk

Commit de53fd7aedb1 ("sched/fair: Fix low cpu usage with high throttling by removing expiration of cpu-local slices") fixes a major performance issue for containerized clouds such as Kubernetes.

Commit de53fd7aedb1 Fixes commit : 512ac999d275 ("sched/fair: Fix bandwidth timer clock drift condition").

Quoted 8 lines
Quoted 5 lines
Quoted 2 lines

This should be applied to all stable kernels that applied commit 512ac999d275, and should probably be applied to all others as well.

As this commit isn't in a released kernel just yet, we should wait to see what happens when it hits people's machines, right?

I think waiting till 5.4 is released would be irresponsible on this one. 512ac999 was recently pushed back into many of the distro

The "released kernel" statement means 5.4-rc1 rather than 5.4, which is just a few days away.

--
Thanks,
Sasha

maintainer
#6 lore raw
Quoted 5 lines · Dave Chiluk

@Greg KH, Qian Cai's compiler warning fix has now been integrated into Linus' tree as commit: 763a9ec06c409

Both de53fd7aedb1 and 763a9ec06c40 are now apart of v5.4-rc1. Can you please queue up these fixes for backport to all stable kernels.

I need an ack from the scheduler maintainers that this is ok to do so...

thanks,

greg k-h

#7 lore raw

@Ben @Ingo @Peter
Can you please please ack this backport request?

Thank you,
Dave Chiluk

Quoted 13 lines · Greg KH
Quoted 5 lines · Dave Chiluk

@Greg KH, Qian Cai's compiler warning fix has now been integrated into Linus' tree as commit: 763a9ec06c409

Both de53fd7aedb1 and 763a9ec06c40 are now apart of v5.4-rc1. Can you please queue up these fixes for backport to all stable kernels.

I need an ack from the scheduler maintainers that this is ok to do so...

thanks,

greg k-h

maintainer
#8 lore raw
Quoted 16 lines · Dave Chiluk

@Ben @Ingo @Peter
Can you please please ack this backport request?

Thank you,
Dave Chiluk

Quoted 9 lines · Greg KH
Quoted 5 lines · Dave Chiluk

@Greg KH, Qian Cai's compiler warning fix has now been integrated into Linus' tree as commit: 763a9ec06c409

Both de53fd7aedb1 and 763a9ec06c40 are now apart of v5.4-rc1. Can you please queue up these fixes for backport to all stable kernels.

I need an ack from the scheduler maintainers that this is ok to do so...

Sure I suppose, but what makes this commit special? Don't you normally take just about anything?

#9 lore raw
Quoted 14 lines · Peter Zijlstra
Quoted 12 lines · Dave Chiluk

@Ben @Ingo @Peter
Can you please please ack this backport request?

Thank you,
Dave Chiluk

Quoted 5 lines · Greg KH
Quoted 3 lines · Dave Chiluk

@Greg KH, Qian Cai's compiler warning fix has now been integrated into Linus' tree as commit: 763a9ec06c409

Quoted 7 lines
Quoted 4 lines
Quoted 4 lines
Quoted 2 lines

Both de53fd7aedb1 and 763a9ec06c40 are now apart of v5.4-rc1. Can you please queue up these fixes for backport to all stable kernels.

I need an ack from the scheduler maintainers that this is ok to do so...

Sure I suppose, but what makes this commit special? Don't you normally take just about anything?

I think this is more a matter of me being a relatively unknown in the scheduler space, and Greg is just being responsible as this looks like a pretty scary fix.

In reality, I probably should have just added "Cc:
[email protected]" to the sign-off area of the initial commit and this conversation wouldn't have been necessary.

maintainer
#10 lore raw
Quoted 23 lines · Dave Chiluk
Quoted 14 lines · Peter Zijlstra
Quoted 12 lines · Dave Chiluk

@Ben @Ingo @Peter
Can you please please ack this backport request?

Thank you,
Dave Chiluk

Quoted 5 lines · Greg KH
Quoted 3 lines · Dave Chiluk

@Greg KH, Qian Cai's compiler warning fix has now been integrated into Linus' tree as commit: 763a9ec06c409

Quoted 7 lines
Quoted 4 lines
Quoted 4 lines
Quoted 2 lines

Both de53fd7aedb1 and 763a9ec06c40 are now apart of v5.4-rc1. Can you please queue up these fixes for backport to all stable kernels.

I need an ack from the scheduler maintainers that this is ok to do so...

Sure I suppose, but what makes this commit special? Don't you normally take just about anything?

Quoted 7 lines

I think this is more a matter of me being a relatively unknown in the scheduler space, and Greg is just being responsible as this looks like a pretty scary fix.

In reality, I probably should have just added "Cc:
[email protected]" to the sign-off area of the initial commit and this conversation wouldn't have been necessary.

That is very true :)

I just tried to apply this to 5.3, 4.19, and 4.14, and it only applies cleanly to 5.3.y. Can you please provide backports for 4.19.y and 4.14.y so that I can queue it up there?

thanks,

greg k-h

#11 lore raw

PATCH 1/2 sched/fair: Fix low cpu usage with high throttling by removing expiration of cpu-local slices

[ Upstream commit de53fd7aedb100f03e5d2231cfce0e4993282425 ]

It has been observed, that highly-threaded, non-cpu-bound applications running under cpu.cfs_quota_us constraints can hit a high percentage of periods throttled while simultaneously not consuming the allocated amount of quota. This use case is typical of user-interactive non-cpu bound applications, such as those running in kubernetes or mesos when run on multiple cpu cores.

This has been root caused to cpu-local run queue being allocated per cpu bandwidth slices, and then not fully using that slice within the period. At which point the slice and quota expires. This expiration of unused slice results in applications not being able to utilize the quota for which they are allocated.

The non-expiration of per-cpu slices was recently fixed by 'commit 512ac999d275 ("sched/fair: Fix bandwidth timer clock drift condition")'. Prior to that it appears that this had been broken since at least 'commit 51f2176d74ac ("sched/fair: Fix unlocked reads of some cfs_b->quota/period")' which was introduced in v3.16-rc1 in 2014. That added the following conditional which resulted in slices never being expired.

if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
	/* extend local deadline, drift is bounded above by 2 ticks */
	cfs_rq->runtime_expires += TICK_NSEC;

Because this was broken for nearly 5 years, and has recently been fixed and is now being noticed by many users running kubernetes
(https://github.com/kubernetes/kubernetes/issues/67577) it is my opinion that the mechanisms around expiring runtime should be removed altogether.

This allows quota already allocated to per-cpu run-queues to live longer than the period boundary. This allows threads on runqueues that do not use much CPU to continue to use their remaining slice over a longer period of time than cpu.cfs_period_us. However, this helps prevent the above condition of hitting throttling while also not fully utilizing your cpu quota.

This theoretically allows a machine to use slightly more than its allotted quota in some periods. This overflow would be bounded by the remaining quota left on each per-cpu runqueueu. This is typically no more than min_cfs_rq_runtime=1ms per cpu. For CPU bound tasks this will change nothing, as they should theoretically fully utilize all of their quota in each period. For user-interactive tasks as described above this provides a much better user/application experience as their cpu utilization will more closely match the amount they requested when they hit throttling. This means that cpu limits no longer strictly apply per period for non-cpu bound applications, but that they are still accurate over longer timeframes.

This greatly improves performance of high-thread-count, non-cpu bound applications with low cfs_quota_us allocation on high-core-count machines. In the case of an artificial testcase (10ms/100ms of quota on 80 CPU machine), this commit resulted in almost 30x performance improvement, while still maintaining correct cpu quota restrictions. That testcase is available at https://github.com/indeedeng/fibtest.

Fixes 512ac999d275 ("sched/fair: Fix bandwidth timer clock drift condition")
Cc Ingo Molnar <[email protected]>
Cc John Hammond <[email protected]>
Cc Jonathan Corbet <[email protected]>
Cc Kyle Anderson <[email protected]>
Cc Gabriel Munos <[email protected]>
Cc Peter Oskolkov <[email protected]>
Cc Cong Wang <[email protected]>
Cc Brendan Gregg <[email protected]>
(cherry picked from commit 25d2df31fd426f34c18227c2cfe2d5e6d14a34fc)
---
 Documentation/scheduler/sched-bwc.txt | 45 ++++++++++++++++++++++
 kernel/sched/fair.c                   | 70 ++++-------------------------------
 kernel/sched/sched.h                  |  4 --
 3 files changed, 52 insertions(+), 67 deletions(-)
Documentation/scheduler/sched-bwc.txt +45 -0
index f6b1873..de583fb 100644
@@ -90,6 +90,51 @@ There are two ways in which a group may become throttled:
 In case b) above, even though the child may have runtime remaining it will not
 be allowed to until the parent's runtime is refreshed.
 
+CFS Bandwidth Quota Caveats
+---------------------------
+Once a slice is assigned to a cpu it does not expire.  However all but 1ms of
+the slice may be returned to the global pool if all threads on that cpu become
+unrunnable. This is configured at compile time by the min_cfs_rq_runtime
+variable. This is a performance tweak that helps prevent added contention on
+the global lock.
+
+The fact that cpu-local slices do not expire results in some interesting corner
+cases that should be understood.
+
+For cgroup cpu constrained applications that are cpu limited this is a
+relatively moot point because they will naturally consume the entirety of their
+quota as well as the entirety of each cpu-local slice in each period. As a
+result it is expected that nr_periods roughly equal nr_throttled, and that
+cpuacct.usage will increase roughly equal to cfs_quota_us in each period.
+
+For highly-threaded, non-cpu bound applications this non-expiration nuance
+allows applications to briefly burst past their quota limits by the amount of
+unused slice on each cpu that the task group is running on (typically at most
+1ms per cpu or as defined by min_cfs_rq_runtime).  This slight burst only
+applies if quota had been assigned to a cpu and then not fully used or returned
+in previous periods. This burst amount will not be transferred between cores.
+As a result, this mechanism still strictly limits the task group to quota
+average usage, albeit over a longer time window than a single period.  This
+also limits the burst ability to no more than 1ms per cpu.  This provides
+better more predictable user experience for highly threaded applications with
+small quota limits on high core count machines. It also eliminates the
+propensity to throttle these applications while simultanously using less than
+quota amounts of cpu. Another way to say this, is that by allowing the unused
+portion of a slice to remain valid across periods we have decreased the
+possibility of wastefully expiring quota on cpu-local silos that don't need a
+full slice's amount of cpu time.
+
+The interaction between cpu-bound and non-cpu-bound-interactive applications
+should also be considered, especially when single core usage hits 100%. If you
+gave each of these applications half of a cpu-core and they both got scheduled
+on the same CPU it is theoretically possible that the non-cpu bound application
+will use up to 1ms additional quota in some periods, thereby preventing the
+cpu-bound application from fully using its quota by that same amount. In these
+instances it will be up to the CFS algorithm (see sched-design-CFS.rst) to
+decide which application is chosen to run, as they will both be runnable and
+have remaining quota. This runtime discrepancy will be made up in the following
+periods when the interactive application idles.
+
 Examples
 --------
 1. Limit a group to 1 CPU worth of runtime.
kernel/sched/fair.c +7 -63

155 lines not on this page · raw message

kernel/sched/sched.h +0 -4
index 452b569..268f560 100644
@@ -280,8 +280,6 @@ struct cfs_bandwidth {
 	ktime_t period;
 	u64 quota, runtime;
 	s64 hierarchical_quota;
-	u64 runtime_expires;
-	int expires_seq;
 
 	short idle, period_active;
 	struct hrtimer period_timer, slack_timer;
@@ -489,8 +487,6 @@ struct cfs_rq {
 
 #ifdef CONFIG_CFS_BANDWIDTH
 	int runtime_enabled;
-	int expires_seq;
-	u64 runtime_expires;
 	s64 runtime_remaining;
 
 	u64 throttled_clock, throttled_clock_task;
Signature · 1 line
1.8.3.1
#12 lore raw

cover Please backport de53fd7aedb1 : sched/fair: Fix low cpu usage with high throttling by removing expiration of cpu-local slices

In reply to Greg KH

Here's the backported patches for 4.14.y. Logic is basically the same, the issue was primarily with patch context. The patches are really back-ports and not cherry-picks because of that, if that's an issue feel free to change the text description.

[PATCH v4.14.y 1/2] sched/fair: Fix low cpu usage with high throttling by [PATCH v4.14.y 2/2] sched/fair: Fix -Wunused-but-set-variable warnings

#13 lore raw

PATCH 1/2 sched/fair: Fix low cpu usage with high throttling by removing expiration of cpu-local slices

[ Upstream commit de53fd7aedb100f03e5d2231cfce0e4993282425 ]

It has been observed, that highly-threaded, non-cpu-bound applications running under cpu.cfs_quota_us constraints can hit a high percentage of periods throttled while simultaneously not consuming the allocated amount of quota. This use case is typical of user-interactive non-cpu bound applications, such as those running in kubernetes or mesos when run on multiple cpu cores.

This has been root caused to cpu-local run queue being allocated per cpu bandwidth slices, and then not fully using that slice within the period. At which point the slice and quota expires. This expiration of unused slice results in applications not being able to utilize the quota for which they are allocated.

The non-expiration of per-cpu slices was recently fixed by 'commit 512ac999d275 ("sched/fair: Fix bandwidth timer clock drift condition")'. Prior to that it appears that this had been broken since at least 'commit 51f2176d74ac ("sched/fair: Fix unlocked reads of some cfs_b->quota/period")' which was introduced in v3.16-rc1 in 2014. That added the following conditional which resulted in slices never being expired.

if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
	/* extend local deadline, drift is bounded above by 2 ticks */
	cfs_rq->runtime_expires += TICK_NSEC;

Because this was broken for nearly 5 years, and has recently been fixed and is now being noticed by many users running kubernetes
(https://github.com/kubernetes/kubernetes/issues/67577) it is my opinion that the mechanisms around expiring runtime should be removed altogether.

This allows quota already allocated to per-cpu run-queues to live longer than the period boundary. This allows threads on runqueues that do not use much CPU to continue to use their remaining slice over a longer period of time than cpu.cfs_period_us. However, this helps prevent the above condition of hitting throttling while also not fully utilizing your cpu quota.

This theoretically allows a machine to use slightly more than its allotted quota in some periods. This overflow would be bounded by the remaining quota left on each per-cpu runqueueu. This is typically no more than min_cfs_rq_runtime=1ms per cpu. For CPU bound tasks this will change nothing, as they should theoretically fully utilize all of their quota in each period. For user-interactive tasks as described above this provides a much better user/application experience as their cpu utilization will more closely match the amount they requested when they hit throttling. This means that cpu limits no longer strictly apply per period for non-cpu bound applications, but that they are still accurate over longer timeframes.

This greatly improves performance of high-thread-count, non-cpu bound applications with low cfs_quota_us allocation on high-core-count machines. In the case of an artificial testcase (10ms/100ms of quota on 80 CPU machine), this commit resulted in almost 30x performance improvement, while still maintaining correct cpu quota restrictions. That testcase is available at https://github.com/indeedeng/fibtest.

Fixes 512ac999d275 ("sched/fair: Fix bandwidth timer clock drift condition")
Cc Ingo Molnar <[email protected]>
Cc John Hammond <[email protected]>
Cc Jonathan Corbet <[email protected]>
Cc Kyle Anderson <[email protected]>
Cc Gabriel Munos <[email protected]>
Cc Peter Oskolkov <[email protected]>
Cc Cong Wang <[email protected]>
Cc Brendan Gregg <[email protected]>
(cherry picked from commit 25d2df31fd426f34c18227c2cfe2d5e6d14a34fc)
---
 Documentation/scheduler/sched-bwc.txt | 45 ++++++++++++++++++++++
 kernel/sched/fair.c                   | 70 ++++-------------------------------
 kernel/sched/sched.h                  |  4 --
 3 files changed, 52 insertions(+), 67 deletions(-)
Documentation/scheduler/sched-bwc.txt +45 -0
index f6b1873..de583fb 100644
@@ -90,6 +90,51 @@ There are two ways in which a group may become throttled:
 In case b) above, even though the child may have runtime remaining it will not
 be allowed to until the parent's runtime is refreshed.
 
+CFS Bandwidth Quota Caveats
+---------------------------
+Once a slice is assigned to a cpu it does not expire.  However all but 1ms of
+the slice may be returned to the global pool if all threads on that cpu become
+unrunnable. This is configured at compile time by the min_cfs_rq_runtime
+variable. This is a performance tweak that helps prevent added contention on
+the global lock.
+
+The fact that cpu-local slices do not expire results in some interesting corner
+cases that should be understood.
+
+For cgroup cpu constrained applications that are cpu limited this is a
+relatively moot point because they will naturally consume the entirety of their
+quota as well as the entirety of each cpu-local slice in each period. As a
+result it is expected that nr_periods roughly equal nr_throttled, and that
+cpuacct.usage will increase roughly equal to cfs_quota_us in each period.
+
+For highly-threaded, non-cpu bound applications this non-expiration nuance
+allows applications to briefly burst past their quota limits by the amount of
+unused slice on each cpu that the task group is running on (typically at most
+1ms per cpu or as defined by min_cfs_rq_runtime).  This slight burst only
+applies if quota had been assigned to a cpu and then not fully used or returned
+in previous periods. This burst amount will not be transferred between cores.
+As a result, this mechanism still strictly limits the task group to quota
+average usage, albeit over a longer time window than a single period.  This
+also limits the burst ability to no more than 1ms per cpu.  This provides
+better more predictable user experience for highly threaded applications with
+small quota limits on high core count machines. It also eliminates the
+propensity to throttle these applications while simultanously using less than
+quota amounts of cpu. Another way to say this, is that by allowing the unused
+portion of a slice to remain valid across periods we have decreased the
+possibility of wastefully expiring quota on cpu-local silos that don't need a
+full slice's amount of cpu time.
+
+The interaction between cpu-bound and non-cpu-bound-interactive applications
+should also be considered, especially when single core usage hits 100%. If you
+gave each of these applications half of a cpu-core and they both got scheduled
+on the same CPU it is theoretically possible that the non-cpu bound application
+will use up to 1ms additional quota in some periods, thereby preventing the
+cpu-bound application from fully using its quota by that same amount. In these
+instances it will be up to the CFS algorithm (see sched-design-CFS.rst) to
+decide which application is chosen to run, as they will both be runnable and
+have remaining quota. This runtime discrepancy will be made up in the following
+periods when the interactive application idles.
+
 Examples
 --------
 1. Limit a group to 1 CPU worth of runtime.
kernel/sched/fair.c +7 -63

155 lines not on this page · raw message

kernel/sched/sched.h +0 -4
index 452b569..268f560 100644
@@ -280,8 +280,6 @@ struct cfs_bandwidth {
 	ktime_t period;
 	u64 quota, runtime;
 	s64 hierarchical_quota;
-	u64 runtime_expires;
-	int expires_seq;
 
 	short idle, period_active;
 	struct hrtimer period_timer, slack_timer;
@@ -489,8 +487,6 @@ struct cfs_rq {
 
 #ifdef CONFIG_CFS_BANDWIDTH
 	int runtime_enabled;
-	int expires_seq;
-	u64 runtime_expires;
 	s64 runtime_remaining;
 
 	u64 throttled_clock, throttled_clock_task;
Signature · 1 line
1.8.3.1
#14 lore raw

PATCH 2/2 sched/fair: Fix -Wunused-but-set-variable warnings

In reply to Dave Chiluk

From: Qian Cai <[email protected]>

[ Upstream commit 763a9ec06c409dcde2a761aac4bb83ff3938e0b3 ]

Commit:

de53fd7aedb1 ("sched/fair: Fix low cpu usage with high throttling by removing expiration of cpu-local slices")

introduced a few compilation warnings:

kernel/sched/fair.c: In function '__refill_cfs_bandwidth_runtime':
kernel/sched/fair.c:4365:6: warning: variable 'now' set but not used [-Wunused-but-set-variable] kernel/sched/fair.c: In function 'start_cfs_bandwidth':
kernel/sched/fair.c:4992:6: warning: variable 'overrun' set but not used [-Wunused-but-set-variable]

Also, __refill_cfs_bandwidth_runtime() does no longer update the expiration time, so fix the comments accordingly.

Cc Linus Torvalds <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Thomas Gleixner <[email protected]>
Fixes de53fd7aedb1 ("sched/fair: Fix low cpu usage with high throttling by removing expiration of cpu-local slices")
---
 kernel/sched/fair.c | 15 +++++----------
 1 file changed, 5 insertions(+), 10 deletions(-)
kernel/sched/fair.c +5 -10
index d5c032e..feeb528 100644
@@ -4091,21 +4091,16 @@ static inline u64 sched_cfs_bandwidth_slice(void)
 }
 
 /*
- * Replenish runtime according to assigned quota and update expiration time.
- * We use sched_clock_cpu directly instead of rq->clock to avoid adding
- * additional synchronization around rq->lock.
+ * Replenish runtime according to assigned quota. We use sched_clock_cpu
+ * directly instead of rq->clock to avoid adding additional synchronization
+ * around rq->lock.
  *
  * requires cfs_b->lock
  */
 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
 {
-	u64 now;
-
-	if (cfs_b->quota == RUNTIME_INF)
-		return;
-
-	now = sched_clock_cpu(smp_processor_id());
-	cfs_b->runtime = cfs_b->quota;
+	if (cfs_b->quota != RUNTIME_INF)
+		cfs_b->runtime = cfs_b->quota;
 }
 
 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
Signature · 1 line
1.8.3.1
maintainer
#15 lore raw
In reply to Dave Chiluk
Quoted 8 lines · Dave Chiluk

Here's the backported patches for 4.14.y. Logic is basically the same, the issue was primarily with patch context. The patches are really back-ports and not cherry-picks because of that, if that's an issue feel free to change the text description.

[PATCH v4.14.y 1/2] sched/fair: Fix low cpu usage with high throttling by [PATCH v4.14.y 2/2] sched/fair: Fix -Wunused-but-set-variable warnings

Both now applied, thanks!

greg k-h

#16 lore raw

cover Please backport de53fd7aedb1 : sched/fair: Fix low cpu usage with high throttling by removing expiration of cpu-local slices

In reply to Greg KH

Here's the backported patches for 4.19.y. Logic is basically the same, the issue was primarily with patch context. The patches are really back-ports and not cherry-picks because of that, if that's an issue feel free to change the text description.

[PATCH v4.19.y 1/2] sched/fair: Fix low cpu usage with high throttling by [PATCH v4.19.y 2/2] sched/fair: Fix -Wunused-but-set-variable warnings

#17 lore raw

PATCH 1/2 sched/fair: Fix low cpu usage with high throttling by removing expiration of cpu-local slices

[ Upstream commit de53fd7aedb100f03e5d2231cfce0e4993282425 ]

It has been observed, that highly-threaded, non-cpu-bound applications running under cpu.cfs_quota_us constraints can hit a high percentage of periods throttled while simultaneously not consuming the allocated amount of quota. This use case is typical of user-interactive non-cpu bound applications, such as those running in kubernetes or mesos when run on multiple cpu cores.

This has been root caused to cpu-local run queue being allocated per cpu bandwidth slices, and then not fully using that slice within the period. At which point the slice and quota expires. This expiration of unused slice results in applications not being able to utilize the quota for which they are allocated.

The non-expiration of per-cpu slices was recently fixed by 'commit 512ac999d275 ("sched/fair: Fix bandwidth timer clock drift condition")'. Prior to that it appears that this had been broken since at least 'commit 51f2176d74ac ("sched/fair: Fix unlocked reads of some cfs_b->quota/period")' which was introduced in v3.16-rc1 in 2014. That added the following conditional which resulted in slices never being expired.

if (cfs_rq->runtime_expires != cfs_b->runtime_expires) {
	/* extend local deadline, drift is bounded above by 2 ticks */
	cfs_rq->runtime_expires += TICK_NSEC;

Because this was broken for nearly 5 years, and has recently been fixed and is now being noticed by many users running kubernetes
(https://github.com/kubernetes/kubernetes/issues/67577) it is my opinion that the mechanisms around expiring runtime should be removed altogether.

This allows quota already allocated to per-cpu run-queues to live longer than the period boundary. This allows threads on runqueues that do not use much CPU to continue to use their remaining slice over a longer period of time than cpu.cfs_period_us. However, this helps prevent the above condition of hitting throttling while also not fully utilizing your cpu quota.

This theoretically allows a machine to use slightly more than its allotted quota in some periods. This overflow would be bounded by the remaining quota left on each per-cpu runqueueu. This is typically no more than min_cfs_rq_runtime=1ms per cpu. For CPU bound tasks this will change nothing, as they should theoretically fully utilize all of their quota in each period. For user-interactive tasks as described above this provides a much better user/application experience as their cpu utilization will more closely match the amount they requested when they hit throttling. This means that cpu limits no longer strictly apply per period for non-cpu bound applications, but that they are still accurate over longer timeframes.

This greatly improves performance of high-thread-count, non-cpu bound applications with low cfs_quota_us allocation on high-core-count machines. In the case of an artificial testcase (10ms/100ms of quota on 80 CPU machine), this commit resulted in almost 30x performance improvement, while still maintaining correct cpu quota restrictions. That testcase is available at https://github.com/indeedeng/fibtest.

Fixes 512ac999d275 ("sched/fair: Fix bandwidth timer clock drift condition")
Cc Ingo Molnar <[email protected]>
Cc John Hammond <[email protected]>
Cc Jonathan Corbet <[email protected]>
Cc Kyle Anderson <[email protected]>
Cc Gabriel Munos <[email protected]>
Cc Peter Oskolkov <[email protected]>
Cc Cong Wang <[email protected]>
Cc Brendan Gregg <[email protected]>
---
 Documentation/scheduler/sched-bwc.txt | 45 ++++++++++++++++++++++
 kernel/sched/fair.c                   | 72 ++++-------------------------------
 kernel/sched/sched.h                  |  4 --
 3 files changed, 52 insertions(+), 69 deletions(-)
Documentation/scheduler/sched-bwc.txt +45 -0
index f6b1873..de583fb 100644
@@ -90,6 +90,51 @@ There are two ways in which a group may become throttled:
 In case b) above, even though the child may have runtime remaining it will not
 be allowed to until the parent's runtime is refreshed.
 
+CFS Bandwidth Quota Caveats
+---------------------------
+Once a slice is assigned to a cpu it does not expire.  However all but 1ms of
+the slice may be returned to the global pool if all threads on that cpu become
+unrunnable. This is configured at compile time by the min_cfs_rq_runtime
+variable. This is a performance tweak that helps prevent added contention on
+the global lock.
+
+The fact that cpu-local slices do not expire results in some interesting corner
+cases that should be understood.
+
+For cgroup cpu constrained applications that are cpu limited this is a
+relatively moot point because they will naturally consume the entirety of their
+quota as well as the entirety of each cpu-local slice in each period. As a
+result it is expected that nr_periods roughly equal nr_throttled, and that
+cpuacct.usage will increase roughly equal to cfs_quota_us in each period.
+
+For highly-threaded, non-cpu bound applications this non-expiration nuance
+allows applications to briefly burst past their quota limits by the amount of
+unused slice on each cpu that the task group is running on (typically at most
+1ms per cpu or as defined by min_cfs_rq_runtime).  This slight burst only
+applies if quota had been assigned to a cpu and then not fully used or returned
+in previous periods. This burst amount will not be transferred between cores.
+As a result, this mechanism still strictly limits the task group to quota
+average usage, albeit over a longer time window than a single period.  This
+also limits the burst ability to no more than 1ms per cpu.  This provides
+better more predictable user experience for highly threaded applications with
+small quota limits on high core count machines. It also eliminates the
+propensity to throttle these applications while simultanously using less than
+quota amounts of cpu. Another way to say this, is that by allowing the unused
+portion of a slice to remain valid across periods we have decreased the
+possibility of wastefully expiring quota on cpu-local silos that don't need a
+full slice's amount of cpu time.
+
+The interaction between cpu-bound and non-cpu-bound-interactive applications
+should also be considered, especially when single core usage hits 100%. If you
+gave each of these applications half of a cpu-core and they both got scheduled
+on the same CPU it is theoretically possible that the non-cpu bound application
+will use up to 1ms additional quota in some periods, thereby preventing the
+cpu-bound application from fully using its quota by that same amount. In these
+instances it will be up to the CFS algorithm (see sched-design-CFS.rst) to
+decide which application is chosen to run, as they will both be runnable and
+have remaining quota. This runtime discrepancy will be made up in the following
+periods when the interactive application idles.
+
 Examples
 --------
 1. Limit a group to 1 CPU worth of runtime.
kernel/sched/fair.c +7 -65

163 lines not on this page · raw message

kernel/sched/sched.h +0 -4
index 9a7c3d0..62058fd 100644
@@ -334,8 +334,6 @@ struct cfs_bandwidth {
 	u64			quota;
 	u64			runtime;
 	s64			hierarchical_quota;
-	u64			runtime_expires;
-	int			expires_seq;
 
 	short			idle;
 	short			period_active;
@@ -555,8 +553,6 @@ struct cfs_rq {
 
 #ifdef CONFIG_CFS_BANDWIDTH
 	int			runtime_enabled;
-	int			expires_seq;
-	u64			runtime_expires;
 	s64			runtime_remaining;
 
 	u64			throttled_clock;
Signature · 1 line
1.8.3.1
#18 lore raw

PATCH 2/2 sched/fair: Fix -Wunused-but-set-variable warnings

In reply to Dave Chiluk

From: Qian Cai <[email protected]>

[ Upstream commit 763a9ec06c409dcde2a761aac4bb83ff3938e0b3 ]

Commit:

de53fd7aedb1 ("sched/fair: Fix low cpu usage with high throttling by removing expiration of cpu-local slices")

introduced a few compilation warnings:

kernel/sched/fair.c: In function '__refill_cfs_bandwidth_runtime':
kernel/sched/fair.c:4365:6: warning: variable 'now' set but not used [-Wunused-but-set-variable] kernel/sched/fair.c: In function 'start_cfs_bandwidth':
kernel/sched/fair.c:4992:6: warning: variable 'overrun' set but not used [-Wunused-but-set-variable]

Also, __refill_cfs_bandwidth_runtime() does no longer update the expiration time, so fix the comments accordingly.

Cc Linus Torvalds <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Thomas Gleixner <[email protected]>
Fixes de53fd7aedb1 ("sched/fair: Fix low cpu usage with high throttling by removing expiration of cpu-local slices")
---
 kernel/sched/fair.c | 19 ++++++-------------
 1 file changed, 6 insertions(+), 13 deletions(-)
kernel/sched/fair.c +6 -13
index cf0f476..e5e8f67 100644
@@ -4305,21 +4305,16 @@ static inline u64 sched_cfs_bandwidth_slice(void)
 }
 
 /*
- * Replenish runtime according to assigned quota and update expiration time.
- * We use sched_clock_cpu directly instead of rq->clock to avoid adding
- * additional synchronization around rq->lock.
+ * Replenish runtime according to assigned quota. We use sched_clock_cpu
+ * directly instead of rq->clock to avoid adding additional synchronization
+ * around rq->lock.
  *
  * requires cfs_b->lock
  */
 void __refill_cfs_bandwidth_runtime(struct cfs_bandwidth *cfs_b)
 {
-	u64 now;
-
-	if (cfs_b->quota == RUNTIME_INF)
-		return;
-
-	now = sched_clock_cpu(smp_processor_id());
-	cfs_b->runtime = cfs_b->quota;
+	if (cfs_b->quota != RUNTIME_INF)
+		cfs_b->runtime = cfs_b->quota;
 }
 
 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
@@ -4924,15 +4919,13 @@ static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
 
 void start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
 {
-	u64 overrun;
-
 	lockdep_assert_held(&cfs_b->lock);
 
 	if (cfs_b->period_active)
 		return;
 
 	cfs_b->period_active = 1;
-	overrun = hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
+	hrtimer_forward_now(&cfs_b->period_timer, cfs_b->period);
 	hrtimer_start_expires(&cfs_b->period_timer, HRTIMER_MODE_ABS_PINNED);
 }
 
Signature · 1 line
1.8.3.1
maintainer
#19 lore raw
In reply to Dave Chiluk
Quoted 7 lines · Dave Chiluk

Here's the backported patches for 4.19.y. Logic is basically the same, the issue was primarily with patch context. The patches are really back-ports and not cherry-picks because of that, if that's an issue feel free to change the text description.

[PATCH v4.19.y 1/2] sched/fair: Fix low cpu usage with high throttling by [PATCH v4.19.y 2/2] sched/fair: Fix -Wunused-but-set-variable warnings

Both now queued up, thanks.

greg k-h