Re: [PATCH v3 0/2] sched/fair: find_energy_efficient_cpu() enhancements

From: Pierre Gondois
Date: Tue May 04 2021 - 05:10:55 EST


Hi Xuewen,

Please let me know if you have some comments,
Pierre

On 5/4/21 10:07 AM, Pierre Gondois wrote:
V2:
- Split the patch in 2. [Quentin]
- Add testing results to the cover-letter. [Dietmar]
- Put back 'rcu_read_unlock()' to unlock the rcu
earlier. [Dietmar]
- Various comments. [Dietmar/Quentin]

V3:
- Layout/phrasing. [Dietmar]

This patchset prevents underflows in find_energy_efficient_cpu().
This is done in the second patch:
sched/fair: Fix negative energy delta in find_energy_efficient_cpu()

The first patch:
sched/fair: Only compute base_energy_pd if necessary
prevents an unnecessary call to compute_energy() if no CPU is available
in a performance domain (pd).
When looping over the pds, it also allows to gather the calls
to compute_energy(), reducing the chances of having utilization signals
being concurrently updated and having a 'negative delta'.

The energy tests of the initial EAS enablement at:
https://lkml.kernel.org/r/20181203095628.11858-1-quentin.perret@xxxxxxx
have been executed using LISA on a Juno-r2 (2xA57 + 4xA53).

To recall the test:
"10 iterations of between 10 and 50 periodic rt-app tasks (16ms period,
5% duty-cycle) for 30 seconds with energy measurement. Unit is Joules.
The goal is to save energy, so lower is better."
"Energy is measured with the onboard energy meter. Numbers include
consumption of big and little CPUs."

+----------+-----------------+-------------------------+
| | Without patches | With patches |
+----------+--------+--------+------------------+------+
| Tasks nb | Mean | CI* | Mean | CI* |
+----------+--------+--------+------------------+------+
| 10 | 6.57 | 0.24 | 6.46 (-1.63%) | 0.27 |
| 20 | 12.44 | 0.21 | 12.44 (-0.01%) | 0.14 |
| 30 | 19.10 | 0.78 | 18.75 (-1.85%) | 0.15 |
| 40 | 27.27 | 0.53 | 27.35 (+0.31%) | 0.33 |
| 50 | 36.55 | 0.42 | 36.28 (-0.74%) | 0.42 |
+----------+-----------------+-------------------------+
CI: confidence interval

For each line, the intervals of values w/ w/o the patches are
overlapping (consider Mean +/- CI). Thus, the energy results shouldn't
have been impacted.

Pierre Gondois (2):
sched/fair: Only compute base_energy_pd if necessary
sched/fair: Fix negative energy delta in find_energy_efficient_cpu()

kernel/sched/fair.c | 68 ++++++++++++++++++++++++++-------------------
1 file changed, 39 insertions(+), 29 deletions(-)