Re: [RFC][PATCH 0/3] cpuidle,teo: Improve TEO tick decisions

[Anna-Maria Behnsen]

Hi,

On Fri, 28 Jul 2023, Peter Zijlstra wrote:

> Hi,
> 
> Wanted to send this yesterday, but my home server died and took everything
> down :/
> 
> These patches are lightly tested but appear to behave as expected.
> 
> 

As I was asked to see if the patches of Raphael improve the behavior, I
rerun the tests with Raphaels v2 as well as with Peters RFC patchset. Here
are the results (compared to upstream):

			upstream		raphael v2		peter RFC

Idle Total		2533	100.00%		1183	100.00%		5563	100.00%
x >= 4ms		1458	57.56%		1151	97.30%		3385	60.85%
4ms> x >= 2ms		91	3.59%		12	1.01%		133	2.39%
2ms > x >= 1ms		56	2.21%		3	0.25%		80	1.44%
1ms > x >= 500us	64	2.53%		1	0.08%		98	1.76%
500us > x >= 250us	73	2.88%		0	0.00%		113	2.03%
250us > x >=100us	76	3.00%		2	0.17%		106	1.91%
100us > x >= 50us	33	1.30%		4	0.34%		75	1.35%
50us > x >= 25us	39	1.54%		4	0.34%		152	2.73%
25us > x >= 10us	199	7.86%		4	0.34%		404	7.26%
10us > x > 5us		156	6.16%		0	0.00%		477	8.57%
5us > x			288	11.37%		2	0.17%		540	9.71%

tick_nohz_next_event()
count			8839790			6142079			36623


Raphaels Approach still does the tick_nohz_get_sleep_length() execution
unconditional. It executes ~5000 times more tick_nohz_next_event() as the
tick is stopped. This relation improved massively in Peters approach
(factor is ~7).

Thanks,

	Anna-Maria