Re: [RFC PATCH 0/6] sched: uclamp sum aggregation

[Dietmar Eggemann]

On 04/10/2023 11:04, Hongyan Xia wrote:
> Current implementation of uclamp leaves many things to be desired.
> There are several problems:
> 
> 1. Max aggregation is fragile. A single task with a high UCLAMP_MAX (or
>    with the default value, which is 1024) can ruin the previous
>    settings the moment it is enqueued, as shown in the uclamp frequency
>    spike problem in Section 5.2 of
>    Documentation/scheduler/sched-util-clamp.rst. Constantly running at
>    1024 utilization implies that the CPU is driven at its max capacity.
>    However, with UCLAMP_MAX, this assumption is no longer true. To
>    mitigate this, one idea is to filter out uclamp values for
>    short-running tasks. However, the effectiveness of this idea remains
>    to be seen.

The difference is that we don't have to lift the uclamp_max cap of
runnable p1's uclamp_max (< 1024) when a short running p2 with
uclamp_max = 1024 becomes runnable? Since now, when this happens, we
would just add p2's util_avg_uclamp to cfs_rq's util_avg_uclamp which is
tiny compared to its uclamp_max = 1024.

> 2. No way to differentiate between UCLAMP_MAX throttled CPU or a CPU
>    running at its peak, as both show utilization of 1024. However, in
>    certain cases it's important to tell the difference, as we still want
>    to take the opportunity to enqueue tasks in the former case.

Is this related to the `best_fits/max_fits` logic in
find_energy_efficient_cpu() (feec()) and select_idle_capacity() (sic())
(commit e5ed0550c04c "sched/fair: unlink misfit task from cpu
overutilized")?
With your approach, having cfs_rq's `util_avg_uclamp` next to its
`util_avg` would allow to see those difference by comparing the two signals?

> It's also debatable whether uclamp should be a frequency hint. An
> example is a system with a mid CPU of capacity 500, and a little CPU
> with capacity 100. If we have 10 tasks with UCLAMP_MIN of 101, then
> util_fits_cpu() will schedule them on the mid CPU because feec()
> thinks they don't fit on the little, even though we should use the
> little core at some point instead of making the mid even more crowded.
> Of course, this in reality doesn't happen because of other mechanisms
> like load balancing, but it's also not good when other mechanisms can

CPU overutilization detection enables CFS load-balance & `sis()->sic()`
and disables feec().

> cancel the effect of uclamp in feec(). A CPU running at capacity 500 for
> 1ms or for 1000ms gives completely different performance levels, so
> trying to fit only the frequency does not give us any performance
> guarantees. If we then talk about not only running at some frequency but
> also for some amount of time, then what we really mean is a capacity
> hint, not a frequency hint.

Isn't CPU frequency and capacity going in the same direction?

 IPC * CPU frequency = Instruction per Second == Performance (Capacity).

And in the scheduler, utilization is the portion of the Capacity
currently used.

What sum aggregation does differently is that you can sum-up
individually clamped utilization contributions and compare them against
capacity rather then being forced to use the maximum value of a clamp
value of one (runnable) task to guide frequency. This avoids those
discontinuity-moments when a task with a high uclamp value switches
between runnable and sleeping state.

> It's even worse when the tasks scheduled on the mid CPU also have
> UCLAMP_MAX values. In the previous example, instead of running at 500, a
> single UCLAMP_MAX, assuming it's 110, can make the mid CPU run at a much
> lower frequency than 500, so it is then a really bad decision to honor
> the UCLAMP_MIN frequency hint and place it on the mid CPU, instead of
> running it on the little CPU which is less crowded, and has pretty much
> the same capacity (100 vs 110) anyway.
> 
> This series attempts to solve these problems by tracking a
> util_avg_uclamp signal in tasks and cfs_rq. At task level,
> p->se.avg.util_avg_uclamp is basically tracking the normal util_avg, but
> clamped within its uclamp min and max. At cfs_rq level, util_avg_uclamp
> must always be the sum of all util_avg_uclamp of all the entities on
> this cfs_rq. As a result, cfs_rq->avg.util_avg_uclamp is the sum
> aggregation of all the clamped values, which indicates the frequency
> this rq should run at and what the utilization is.
> 
> This idea solves Problem 1 by capping the utilization of an
> always-running task throttled by UCLAMP_MAX. Although the task (denoted
> by Task A) has no idle time, the util_avg_uclamp signal gives its
> UCLAMP_MAX value instead of 1024, so even if another task (Task B) with
> a UCLAMP_MAX value at 1024 joins the rq, the util_avg_uclamp is A's
> UCLAMP_MAX plus B's utilization, instead of 1024 plus B's utilization,
> which means we no longer have the frequency spike problem caused by B.
> This should mean that we might completely avoid the need for uclamp
> filtering.

That would be very nice since I remember that this filtering approach
hat to figure out the actual runtime of the task and the implemention
couldn't be just in the sched class code but had to be done in core code
as well.

> 
> It also solves Problem 2 by tracking the capped utilization of a rq.
> Using util_avg_uclamp, we are able to differentiate between a CPU
> throttled by UCLAMP_MAX and one that is actually running at its peak
> capacity. An always-running rq with a task having UCLAMP_MAX at 300 will
> report a cfs_rq util_avg_uclamp at 300, not 1024, which means task
> placement sees spare capacity on this CPU and will allocate some tasks
> to it, instead of treating it the same way as a CPU actually running at
> the peak. This brings us closer to the benefits of Android's sum
> aggregation (see code related to CONFIG_USE_GROUP_THROTTLE at
> https://android.googlesource.com/kernel/gs/+/refs/heads/android-gs-raviole-5.10-android12-d1/drivers/soc/google/vh/kernel/sched/fair.c#510),

That's true although the latest Pixel phone does not seem to rely on
CONFIG_USE_GROUP_THROTTLE anymore but on customized PELT tracking for so
called Vendor CFS Util Groups and its sum aggregation on CPU (root
cfs_rq) level(CONFIG_USE_VENDOR_GROUP_UTIL).

CONFIG_USE_GROUP_THROTTLE was using the fact that Android only uses 1.
level taskgroups to map its different task types (e.g. foreground,
background etc.) into:

  for_each_leaf_cfs_rq_safe()
    calc subgroup_util_sum and throttled_subgroup_util_sum and skip root

  cpu_util = root_util - subgroup_util_sum + throttled_subgroup_util_sum

So summing up the (throttled (capped)) contritions of the 1. level
taskgroups plus the root taskgroup util.

With CONFIG_USE_VENDOR_GROUP_UTIL each tasks has a vendor pointer into
an CFS util array indexed by an Android task type ID to which the task
contributes its util. CPU util is then sum aggregated over this array.

We should recall that this is all done because the current uclamp-max
max aggression isn't working for Androids use-cases.

So to overcome this issue in mainline is key here.

> which shows energy benefits because we are able to schedule tasks on
> smaller cores which are UCLAMP_MAX capped, instead of finding a fresh
> big CPU. However, a major difference is that this series has an
> amortized cost of O(1), instead of O(n) in cpu_util_next() in Android
> kernels.
> 
> It avoids the shortcomings from uclamp-being-a-frequency-hint. Under sum
> aggregation, the scheduler sees the sum aggregated util_avg_uclamp and
> will avoid the problem of enqueueing too many tasks just to fit the
> frequency, and will place tasks on little CPUs in the previous example.
> 
> Patch 2/6 tries to simulate PELT decay in the new util_avg_uclamp
> signal, as this gives us gradual decay of utilization which avoids
> premature CPU frequency drops. This is a major caveat of this series. We
> should explore if there's a better way to integrate uclamp directly into
> the util_avg signal, instead of introducing a new util_avg_uclamp and
> then simulate PELT on it.
> 
> This new design significantly simplifies uclamp logic.
> Patch 4/6 removes the tri-state return value of util_fits_cpu().
> Patch 5/6 then completely removes uclamp buckets. Because the
> util_avg_uclamp is already a clamped value propagated from bottom to
> top, there's no need to clamp anything at the top level and we can just
> use this value for frequency selection and spare capacity search.
> Basically, the idea is that all uclamp logic happens inside
> util_avg_uclamp, and other code using this new signal can just use it
> like util_avg, without even knowing that uclamp exists. This drastically
> reduces the complexity of uclamp and makes the code considering all the
> uclamp corner cases unnecessary. At the end of the series, we remove 749
> lines of code while adding a bit more than 300 (although once we update
> Documentation.rst, it will be a bit more than that).
> 
> Note that this series is still considered RFC status. TODO items are:
> 
> 1. Implement sum aggregation for RT tasks.
> 2. Improve handling of cpu_util(boost).

What about the integration with util_est here?

In cpu_util(), &rq->cfs->avg.util_avg is replaced by
rq->root_cfs_util_uclamp

and in

task_util_est() (should be actually named task_util() to be in sync with
cpu_util(), i.e. returning max(util, util_est)), task_util(p) returns
p->se.avg.util_avg_uclamp.

Are there use cases for the original avg.util_avg still in place?

> TESTING:
> 
> Initial test and benchmark results suggest that sum aggregation not only
> is significantly simpler, but generally performs much better in
> achieving what uclamp is supposed to do. Two notebooks are shared at
> 
> https://nbviewer.org/github/honxia02/sum_aggre_notebooks/blob/main/upstream.ipynb
> https://nbviewer.org/github/honxia02/sum_aggre_notebooks/blob/main/sum.ipynb
> 
> The experiments done in notebooks are on Arm Juno r2 board. CPU0-3 are
> little cores with capacity of 383. CPU4-5 are big cores. The rt-app
> profiles used for these experiments are included in the notebooks.
> 
> Scenario 1: Scheduling 4 always-running tasks with UCLAMP_MAX at 200.
> 
> The scheduling decisions are plotted in Out[11] and Out[14]
> respectively. Max aggregation fails to recognize the opportunity to run
> all of them on the efficient little Power Domain (PD), whereas sum
> aggregation runs all 4 on the little PD, leaving the big PD completely
> powered off, satisfying uclamp requests while saving power.

Does `upstream` already contain the v6.7 fixes `Fix uclamp code corner
cases` ?

https://lkml.kernel.org/r/20230916232955.2099394-1-qyousef@xxxxxxxxxxx

> Scenario 2: Scheduling 4 tasks with UCLAMP_MIN and UCLAMP_MAX at a value
> slightly above the capacity of the little CPU.
> 
> Results are in Out[17] and Out[82]. The purpose is to use UCLAMP_MIN to
> place tasks on the big core. Max aggregation is pretty much in an
> overutilized state the entire time. Sum aggregation sees that the big
> CPU can hold two such tasks (satisfying both UCLAMP_MIN and UCLAMP_MAX
> requests for all 4 tasks) on each big CPU and quickly settles down, and
> is still under EAS without overutilization.

thread0-[0-3] uclamp_max = 309. So p->se.avg.util_avg_uclamp is
constrained by this value for all 4 tasks, letting 2 tasks fit on each
of the big CPUs. You have to zoom in into Out[82] to actually see this.

And I guess for max aggregation cpu_overutilized() can't hold the clamp
continuously because of all the other short running uclamp_max = 1024
(default) tasks on the rq's.

> Scenario 3: Task A is a task with a small utilization pinned to CPU4.
> Task B is an always-running task pinned to CPU5, but UCLAMP_MAX capped
> at 200. After a while, task A is then pinned to CPU5, joining B.
> 
> Results are in Out[23] and Out[95]. The blue util_avg signal is the
> original root CFS util_avg. The yellow line is the root CFS utilization
> after considering uclamp. Max aggregation sees a frequency
> spike at 579.1s. When zoomed in, one can see square-wave-like
> utilization values because of A periodically going to sleep. When A
> wakes up, its default UCLAMP_MAX of 1024 will uncap B and reach the
> highest CPU frequency. When A sleeps, B's UCLAMP_MAX will be in effect
> and will reduce rq utilization to 200. This happens repeatedly, hence
> the square wave. In contrast, sum aggregation sees a normal increase in
> utilization when A joins B, at around 2507s, without any square-wave
> behavior.

Makes sense. But there shouldn't be a root_cfs_util_uclamp in main?
Which signal does the yellow line represent in Out[23]?

> Scenario 4: 4 always-running tasks with UCLAMP_MAX of 120 pinned to the
> little PD (CPU0-3). 4 same tasks pinned to the big PD (CPU4-5).
> After a while, remove the CPU pinning of the 4 tasks on the big PD.
> 
> Results are in Out[29] and Out[101]. Max aggregation fails to identify
> that all 8 tasks can be packed on the little PD, whereas sum
> aggregation immediately moves the 4 tasks pinned to big PD to the
> little PD the moment pinning is removed. Sum aggregation understands
> that even when the rq seems to have utilization of 1024, this is because
> of UCLAMP_MAX and there's still opportunity to pack 2 such tasks on each
> little CPU, leaving the big PD untouched, saving power.

Looks good to me.

> 
> BENCHMARKS:
> 
> One TODO item is to handle cpu_util(boost) better. The current handling
> of boost is far from ideal and is a known caveat. All below benchmarks
> numbers are done without any boosting in cpu_util(), in both max and sum
> aggregation.

Maybe to early for black-box tests at this stage but good to see that
nothing seems to go sideways with uclamp sum aggregation.


> 
> Geekbench 6 (on Rock-5B board)
> +-----+-------------+------------+
> |     | Single-core | Multi-core |
> +-----+-------------+------------+
> | Max |      800.6  |     2977.0 |
> | Sum |      801.2  |     2981.4 |
> +-----+-------------+------------+
> 
> No regression is seen after switching to sum aggregation.
> 
> Jankbench (backporting sum aggregation to Android 5.18 mainline kernel)
> 
> +------+-----------------+-------+-----------+
> |      |    variable     | value | perc_diff |
> +------+-----------------+-------+-----------+
> | main | jank_percentage |  1.1  |    0.00%  |
> |  sum | jank_percentage |  0.5  |  -53.92%  |
> +------+-----------------+-------+-----------+
> 
> +------------+--------+------+-------+-----------+
> |            | metric |  tag | value | perc_diff |
> +------------+--------+------+-------+-----------+
> |    CPU     | gmean  | main | 166.1 |   0.00%   |
> |  CPU-Big   | gmean  | main | 55.1  |   0.00%   |
> | CPU-Little | gmean  | main | 91.7  |   0.00%   |
> |  CPU-Mid   | gmean  | main | 19.2  |   0.00%   |
> |    CPU     | gmean  |  sum | 161.3 |  -2.85%   |
> |  CPU-Big   | gmean  |  sum | 52.9  |  -3.97%   |
> | CPU-Little | gmean  |  sum | 86.7  |  -5.39%   |
> |  CPU-Mid   | gmean  |  sum | 21.6  |  12.63%   |
> +------------+--------+------+-------+-----------+
> 
> The TL;DR of Jankbench is that sum aggregation reduces jank by 54% while
> with 2.85% less CPU power. Note that this benchmark on Pixel 6 by
> default has some sort of uclamp feedback applied to it. When jank is
> detected, certain display and benchmark threads are applied with a
> UCLAMP_MIN value of 512 (any help in identifying where these UCLAMP_MIN
> values come from in the mainline Android kernel is appreciated). In
> contrast, we constantly apply a UCLAMP_MIN value of 60 to these threads
> without any feedback loop. If a similar feedback loop is involved to
> only apply UCLAMP_MIN when needed, power consumption can be expected to
> be even lower.

[...]