Re: [PATCH] cpuidle: menu: find the typical interval by a heuristic classification method

[Christian Loehle <christian.loehle@arm.com>]

On 7/17/25 11:11, 朱恺乾 wrote:
> The iterations of deviation calculation gives too less predictions on
> the idle interval by trying to find a single repeating pattern from the
> whole history. This is not always the case when the workload is flowing.
> 
> This algorithm assumes there're multiple repeating patterns heuristically,
> and tries to determine which is the most promising one from the averages
> of different idle states. It also takes the occurrence sequence into
> consideration, and gives the prediction close to the recent idle.
> 
> This increased the shallow idle states detected, but the difference in deep
> sleep time didn't change a lot. The performance on my platform, as
> expected, has improved.
> 
> Before:
> Multi-Core Score              7279
> Overall    above   under
>   34107    0.00    2.75
>    8200   59.90    7.02
>   29881   57.06    0.00
> 
> After:
> Multi-Core Score              7365
> Overall    above   under
>   49913    0.00    6.43
>    7881   44.51   18.08
>   23108   52.38    0.00
> 
> There's another re-classification method, which, instead of looking for the
> repeating-interval, tends to find the repeating state. It gives a better result
> on performance gain, but may hurt the power consumption.
> 
> if (best_state == drv->state_count - 1 || state_avg[best_state] == 0) {
> adj_weight[best_state] += weights[i];
> adj_avg[best_state] += value;
> adj_hit[best_state]++;
> } else if (best_diff < state_avg[best_state] * 2) {
> adj_weight[best_state] += weights[i];
> adj_avg[best_state] += value;
> adj_hit[best_state]++;
> } else {
> adj_weight[best_state + 1] += weights[i];
> adj_avg[best_state + 1] += value;
> adj_hit[best_state + 1]++;
> }
> 
> Repeating State:
> Multi-Core Score              7421
> Overall    above   under
>   60857    0.00    8.30
>    3838   29.88   18.42
>   15318   39.05    0.00
> 
> 

Would be nice to have statistical significance on those (I'm assuming Geekbench)
scores and a description of the systems idle state and maybe a comparison to
shallowest (only WFI enabled) and deepest (only max idle state enabled).