Re: [RFT][PATCH v1 0/5] cpuidle: menu: Avoid discarding useful information when processing recent idle intervals

[Aboorva Devarajan <aboorvad@linux.ibm.com>]

On Mon, 2025-02-24 at 11:57 +0530, Aboorva Devarajan wrote:
> On Thu, 2025-02-06 at 15:21 +0100, Rafael J. Wysocki wrote:
> 
> 
> 
> 
> So for the entire series:
> 
> Tested-by: Aboorva Devarajan <aboorvad@linux.ibm.com>
> 
> I'm also trying a minimal unit fuzz-test with the pre- and post- patched version of the get_typical_interval 
> function to understand this better, will post the results soon.
> 
> 
In addition to the previous tests, I also reviewed and tested how get_typical_interval
predicts the idle duration before and after the patch by mimicking the function in
userspace for better unit fuzz testing [1].

With the patch get_typical_interval function now produces more predictable values, whereas
in the previous implementation it frequently returned UINT_MAX. The behavior with the patch
seems to be more reasonable compared to before.  

Here are the test results  

1. Low Variance:

When the history of CPU idle durations (8 intervals) is relatively uniform with low variance,
the predicted idle duration is unchanged between the patched and unpatched versions:  

| Test Case | Intervals                                    | Before | After | Difference |
|-----------|----------------------------------------------|--------|-------|------------|
| 1         | 100,105,110,115,120,125,130,135              | 117    | 117   | 0          |
| 2         | 200,205,210,215,220,225,230,235              | 217    | 217   | 0          |
| 3         | 500,505,510,515,520,525,530,535              | 517    | 517   | 0          |
| 4         | 1000,1005,1010,1015,1020,1025,1030,1035      | 1017   | 1017  | 0          |

 2. High Variance
 
For cases with high variance, the nonpatched function returned UINT_MAX,

After the patch, the function now returns reasonable values:  

| Test Case | Intervals                                  | Before      | After | Difference       |
|-----------|--------------------------------------------|-------------|-------|------------------|
| 5         | 99,198,297,396,495,594,693,792             | 4294967295  | 594   | -4294966701      |
| 6         | 1000,2000,3000,4000,5000,6000,7000,8000    | 4294967295  | 6000  | -4294961295      |
| 7         | 40,140,240,340,440,540,640,740             | 4294967295  | 540   | -4294966755      |
| 8         | 90,590,1090,1590,2090,2590,3090,3590       | 4294967295  | 2590  | -4294964705      |
| 9         | 42,142,242,342,442,542,642,742             | 4294967295  | 542   | -4294966753      |
| 10        | 70,570,1070,1570,2070,2570,3070,3570       | 4294967295  | 2570  | -4294964725      |
| 11        | 44,144,244,344,444,544,644,744             | 4294967295  | 544   | -4294966751      |

 3. Low-end Outliers 

The patch removes variance from low-end values as well,
Without the patch, the function only filtered high-end outliers, but now it correctly eliminates both
high and low ends.

| Test Case | Intervals                                 | Before      | After | Difference  |
|-----------|-------------------------------------------|-------------|-------|-------------|
| 12        | 1,200,200,250,250,230,220,260             | 4294967295  | 230   | -4294967065 |
| 13        | 100000,200,200,250,250,230,220,260        | 230         | 230   | 0           |


 4. As far as I understand, the function only returns UINT_MAX when all values are 0, negative, or the
computed average itself is UINT_MAX.  

| Test Case | Intervals                                   | Before      | After       | Difference |
|-----------|---------------------------------------------|-------------|-------------|------------|
| 14        | 4294967295,4294967295,4294967295,4294967295 | 4294967295  | 4294967295  | 0          |
| 15        | 0,0,0,0,0,0,0,0                             | 4294967295  | 4294967295  | 0          |

The updated get_typical_interval function avoids unnecessary UINT_MAX returns, handles both low and high end
outliers, and gives more reliable predictions in high-variance cases. It only returns UINT_MAX when absolutely
necessary, and this will help in not selecting deeper idle state unless it is needed. So, I'm good with
the patch. 

Refer [2] for more test results.

[1] - https://github.com/AboorvaDevarajan/linux-utils/blob/main/cpuidle/cpuidle-predict-duration/predict_cpuidle_interval.c
[2] - https://github.com/AboorvaDevarajan/linux-utils/blob/main/cpuidle/cpuidle-predict-duration/out.predict.csv


Thanks,
Aboorva