Skip to content

cpuidle/menu: Address performance drop from favoring physical over polling cpuidle state

Aboorva Devarajan linux-pm lkml 1 version In review

Get series with b4
b4 am [email protected]

Versions newest first

  1. v1 Latest1 patch 16 repliesIn review

Patches in v1 No applies recorded yet

#Subject and filesSizeRepliesApplied
1/1cpuidle/menu: avoid prioritizing physical state over polling statedrivers/cpuidle/governors/menu.c+0 -113 replies

Reviews trailers per patch

No Reviewed-by, Acked-by or Tested-by trailers for v1 yet.

Read the v1 thread

Cover letter [PATCH 0/1] cpuidle/menu: Address performance drop from favoring physical over polling cpuidle state

This patch aims to discuss a potential performance degradation that can occur in certain workloads when the menu governor prioritizes selecting a physical idle state over a polling state for short idle durations.

Note: This patch is intended to showcase a performance degradation, applying this patch could lead to increased power consumption due to the trade-off between performance and power efficiency, potentially causing a higher preference for performance at the expense of power usage.

==================================================
System details in which the degradation is observed:

$ uname -r
6.10.0+
$ lscpu
Architecture:             ppc64le
  Byte Order:             Little Endian
CPU(s):                   160
  On-line CPU(s) list:    0-159
Model name:               POWER10 (architected), altivec supported
  Model:                  2.0 (pvr 0080 0200)
  Thread(s) per core:     8
  Core(s) per socket:     3
  Socket(s):              6
  Physical sockets:       4
  Physical chips:         2
  Physical cores/chip:    6
Virtualization features:
  Hypervisor vendor:      pHyp
  Virtualization type:    para
Caches (sum of all):
  L1d:                    1.3 MiB (40 instances)
  L1i:                    1.9 MiB (40 instances)
  L2:                     40 MiB (40 instances)
  L3:                     160 MiB (40 instances)
NUMA:
  NUMA node(s):           6
  NUMA node0 CPU(s):      0-31
  NUMA node1 CPU(s):      32-71
  NUMA node2 CPU(s):      72-79
  NUMA node3 CPU(s):      80-87
  NUMA node4 CPU(s):      88-119
  NUMA node5 CPU(s):      120-159
$ cpupower idle-info
CPUidle driver: pseries_idle
CPUidle governor: menu
analyzing CPU 0:

Number of idle states: 2
Available idle states: snooze CEDE
snooze:
Flags/Description: snooze
Latency: 0
Residency: 0
Usage: 6229
Duration: 402142
CEDE:
Flags/Description: CEDE
Latency: 12
Residency: 120
Usage: 191411
Duration: 36329999037

==================================================

The menu governor contains a condition that selects physical idle states over, such as the CEDE state over polling state, by checking if their exit latency meets the latency requirements. This can lead to performance drops in workloads with frequent short idle periods.

The specific condition which causes degradation is as below (menu governor):

```
if (s->target_residency_ns > predicted_ns) {
    ...
    if ((drv->states[idx].flags & CPUIDLE_FLAG_POLLING) &&
        s->exit_latency_ns <= latency_req &&
        s->target_residency_ns <= data->next_timer_ns) {
        predicted_ns = s->target_residency_ns;
        idx = i;
        break;
    }
    ...
}
```

This condition can cause the menu governor to choose the CEDE state on Power Systems (residency: 120 us, exit latency: 12 us) over a polling state, even when the predicted idle duration is much shorter than the target residency of the physical state. This misprediction leads to performance degradation in certain workloads.

==================================================
Test Results
==================================================

This issue can be clearly observed with the below test.

A test with multiple wakee threads and a single waker thread was run to demonstrate this issue. The waker thread periodically wakes up the wakee threads after a specific sleep duration, creating a repeating of sleep -> wake pattern. The test was run for a stipulated period, and cpuidle statistics are collected.

./cpuidle-test -a 0 -b 10 -b 20 -b 30 -b 40 -b 50 -b 60 -b 70 -r 20 -t 60

==================================================
Results (Baseline Kernel):
==================================================
Wakee 0[PID 8295] affined to CPUs: 10,
Wakee 2[PID 8297] affined to CPUs: 30,
Wakee 3[PID 8298] affined to CPUs: 40,
Wakee 1[PID 8296] affined to CPUs: 20,
Wakee 4[PID 8299] affined to CPUs: 50,
Wakee 5[PID 8300] affined to CPUs: 60,
Wakee 6[PID 8301] affined to CPUs: 70,
Waker[PID 8302] affined to CPUs: 0,

|-----------------------------------|-------------------------|-----------------------------|
| Metric                            | snooze                  | CEDE                        |
|-----------------------------------|-------------------------|-----------------------------|
| Usage                             | 47815                   | 2030160                     |
| Above                             | 0                       | 2030043                     |
| Below                             | 0                       | 0                           |
| Time Spent (us)                   | 976317 (1.63%)          | 51046474 (85.08%)           |
| Overall average sleep duration    | 28.721 us               |                             |
| Overall average wakeup latency    | 6.858 us                |                             |
|-----------------------------------|-------------------------|-----------------------------|

In this test, without the patch, the CPU often enters the CEDE state for sleep durations of around 20-30 microseconds, even though the CEDE state's residency time is 120 microseconds. This happens because the menu governor prioritizes the physical idle state (CEDE) if its exit latency is within the latency limits. It also uses next_timer_ns for comparison, which can be farther off than the actual idle duration as it is more predictable, instead of using predicted idle duration as a comparision point with the target residency.

Inference: The "Above" metric in the table counts instances where the state was entered but the observed idle duration was too short. The data shows that the CEDE state's "Above" metric is high, with around 99.9% of the CEDE state usage exceeding the expected cpuidle state.

==================================================
Results (With Patch):
==================================================
Wakee 0[PID 8295] affined to CPUs: 10,
Wakee 2[PID 8297] affined to CPUs: 30,
Wakee 3[PID 8298] affined to CPUs: 40,
Wakee 1[PID 8296] affined to CPUs: 20,
Wakee 4[PID 8299] affined to CPUs: 50,
Wakee 5[PID 8300] affined to CPUs: 60,
Wakee 6[PID 8301] affined to CPUs: 70,
Waker[PID 8302] affined to CPUs: 0,

|-----------------------------------|-------------------------|-----------------------------|
| Metric                            | snooze                  | CEDE                        |
|-----------------------------------|-------------------------|-----------------------------|
| Usage                             | 1853343                 | 6                           |
| Above                             | 0                       | 6                           |
| Below                             | 2                       | 0                           |
| Time Spent (us)                   | 51055087 (85.09%)       | 356 (0.00%)                 |
| Overall average sleep duration    | 32.268 us               |                             |
| Overall average wakeup latency    | 4.382 us                |                             |
|-----------------------------------|-------------------------|-----------------------------|

When the condition that prioritizes the selection of physical idle states based on exit latency is removed, these metrics improve significantly. The reduction in "above" metric for the CEDE state shows that the governor’s selection of cpuidle state are now more accurate that before, leading to better overall performance.

==================================================
Performance Improvements:
==================================================
The patch improved the performance of OLTP benchmarks by around 5-10%.

The following table summarizes the performance improvements observed in multiple OLTP query tests:

|----|----------|------------------|---------------------------|
| #  | Base (s) | Base + Patch (s) | % Performance Improvement |
|    | Kernel   | Kernel           |    (lower the better)     |
|----|----------|------------------|---------------------------|
| 1  | 15.77    | 14.14            | -10.34%                   |
| 2  | 0.87     | 0.86             | -0.90%                    |
| 3  | 4.05     | 3.78             | -6.70%                    |
| 4  | 0.61     | 0.59             | -3.10%                    |
| 5  | 1.18     | 1.14             | -3.47%                    |
| 6  | 4.11     | 3.84             | -6.59%                    |
| 7  | 0.75     | 0.75             | -0.24%                    |
| 8  | 14.91    | 13.56            | -9.03%                    |
| 9  | 4.02     | 3.86             | -3.90%                    |
| 10 | 14.93    | 13.65            | -8.57%                    |
| 11 | 15.62    | 14.03            | -10.20%                   |
| 12 | 2.15     | 2.05             | -4.52%                    |
| 13 | 7.92     | 7.35             | -7.29%                    |
| 14 | 4.08     | 4.07             | -0.08%                    |
| 15 | 2.28     | 2.10             | -8.12%                    |
| 16 | 1.14     | 1.11             | -3.09%                    |
| 17 | 2.14     | 2.06             | -3.64%                    |
| 18 | 2.54     | 2.36             | -7.16%                    |
| 19 | 3.72     | 3.68             | -1.05%                    |
| 20 | 2.12     | 2.05             | -3.28%                    |
| 21 | 8.17     | 7.46             | -8.69%                    |
| 22 | 3.98     | 3.83             | -3.58%                    |
| 23 | 2.45     | 2.30             | -6.12%                    |
| 24 | 2.13     | 2.06             | -3.61%                    |
| 25 | 15.05    | 13.74            | -8.70%                    |
| 26 | 4.49     | 4.16             | -7.43%                    |
| 27 | 3.75     | 3.66             | -2.37%                    |
| 28 | 2.90     | 2.87             | -0.96%                    |
| 29 | 4.37     | 4.03             | -7.83%                    |
| 30 | 2.57     | 2.43             | -5.19%                    |
| 31 | 2.78     | 2.71             | -2.30%                    |
| 32 | 3.94     | 3.78             | -4.27%                    |
| 33 | 0.70     | 0.69             | -1.48%                    |
| 34 | 2.15     | 2.08             | -3.62%                    |
| 35 | 8.30     | 7.52             | -9.38%                    |
| 36 | 2.15     | 2.08             | -3.25%                    |
| 37 | 4.37     | 4.03             | -7.88%                    |
| 38 | 4.16     | 3.88             | -6.78%                    |
| 39 | 1.12     | 1.12             | -0.02%                    |
| 40 | 1.22     | 1.17             | -4.19%                    |
| 41 | 8.45     | 7.67             | -9.27%                    |
| 42 | 7.91     | 7.31             | -7.51%                    |
| 43 | 1.12     | 1.12             | -0.08%                    |
| 44 | 0.68     | 0.65             | -3.86%                    |
| 45 | 2.12     | 2.00             | -5.47%                    |
| 46 | 3.11     | 3.04             | -2.36%                    |
| 47 | 2.29     | 2.13             | -6.88%                    |
| 48 | 8.37     | 7.61             | -9.02%                    |
| 49 | 2.61     | 2.46             | -5.75%                    |
| 50 | 7.92     | 7.30             | -7.79%                    |
| 51 | 2.85     | 2.77             | -2.58%                    |
| 52 | 0.79     | 0.78             | -0.87%                    |
| 53 | 2.90     | 2.79             | -3.63%                    |
| 54 | 8.03     | 7.40             | -7.87%                    |
| 55 | 2.14     | 2.00             | -6.53%                    |
| 56 | 15.54    | 13.93            | -10.38%                   |
| 57 | 8.07     | 7.45             | -7.62%                    |
| 58 | 8.48     | 7.70             | -9.20%                    |
| 59 | 1.08     | 1.04             | -3.80%                    |
| 60 | 1.14     | 1.13             | -0.91%                    |
| 61 | 4.28     | 3.95             | -7.84%                    |
| 62 | 8.40     | 7.62             | -9.36%                    |
| 63 | 7.91     | 7.38             | -6.78%                    |
| 64 | 2.16     | 2.10             | -2.92%                    |
| 65 | 7.78     | 7.18             | -7.60%                    |
| 66 | 4.22     | 4.19             | -0.84%                    |
| 67 | 2.99     | 2.88             | -3.61%                    |
| 68 | 8.44     | 7.56             | -10.33%                   |
| 69 | 4.10     | 3.74             | -8.89%                    |
| 70 | 4.25     | 3.98             | -6.37%                    |
| 71 | 0.61     | 0.61             | -0.39%                    |
| 72 | 3.89     | 3.56             | -8.39%                    |
|----|----------|------------------|---------------------------|

==================================================
Observation:
==================================================

The condition under discussion is intended to prioritize physical idle state when their exit latency is within the acceptable limits thereby improving the power efficiency and overall SMT folding. However, in systems where the target residency value is significantly higher than the exit latency, this leads to frequent selection of states like CEDE in Power Systems, which is less efficient for short sleep durations.

Whereas, in systems where the residency value and exit latency are almost equal, this condition might not adversely affect the performance.

In general scenarios, there seems to be a preference for prioritizing energy efficiency, as shown by the commit message 0c313cb (“cpuidle: menu:
Fallback to polling if the next timer event is near”). This change involved reverting to next_timer_ns as a comparison point instead of using the predicted idle duration. While next_timer_ns is more precise, it might not account for very short idle durations, which could affect performance in certain situations, even though it helps reduce energy consumption.

But, any feedback and suggestions on better improving the power-performance tradeoff in menu governor to address the performance issue in specific cases as mentioned above will be very helpful.

Thanks
Aboorva

Aboorva Devarajan (1):
cpuidle/menu: avoid prioritizing physical state over polling state

 drivers/cpuidle/governors/menu.c | 11 -----------
 1 file changed, 11 deletions(-)
Signature · 1 line
2.39.3