Re: [RFC 2/2] sched/fair: prefer a CPU in the "lowest" idle state

[Namhyung Kim <namhyung@kernel.org>]

Hi Michael,

On Thu, 31 Jan 2013 16:57:49 +0800, Michael Wang wrote:
> On 01/31/2013 04:45 PM, Michael Wang wrote:
>> On 01/31/2013 04:24 PM, Michael Wang wrote:
>>> On 01/31/2013 03:40 PM, Namhyung Kim wrote:
>>>> On Thu, 31 Jan 2013 15:30:02 +0800, Michael Wang wrote:
>>>>> On 01/31/2013 02:58 PM, Namhyung Kim wrote:
>>>>>> But AFAIK the number of states in cpuidle is usually less than 10 so maybe
>>>>>> we can change the weight then, but there's no promise...
>>>>>
>>>>> And I just got another case we should take care:
>>>>>
>>>>> 	group 0		cpu 0			cpu 1
>>>>> 			power index 8		power index 8
>>>>>
>>>>>
>>>>> 	group 1		cpu 2			cpu 3
>>>>> 			power index 0		load 15
>>>>>
>>>>> so load of group 0 is 16 and group 1 is 15, but group 0 is better...
>>>>
>>>> Maybe it's not.  The cpus in group 0 are in a lower power state so that
>>>> there will be a benefit to select cpu 2 from the power' PoV IMHO.  Also
>>>> such a low power state has a longer exit latency so that we should
>>>> choose cpu2 to get a better performance and it's the basic idea of this
>>>> patchset I believe.
>>>
>>> Well, this case is just to notify that, we may face the comparison
>>> between load and index, not between index and index, I just doubt there
>>> won't be a rule which could take care both, besides, comparison between
>>> load and index is strange...
>> 
>> Oh, I miss the point that you call it 'idle load', hmm...may be it could
>> works, if we could scale the current load number, then we will have more
>> 'space' for 'idle load'.
>
> And some tips here:
>
> /*      
>  * Increase resolution of nice-level calculations for 64-bit architectures.
>  * The extra resolution improves shares distribution and load balancing of
>  * low-weight task groups (eg. nice +19 on an autogroup), deeper taskgroup
>  * hierarchies, especially on larger systems. This is not a user-visible change
>  * and does not change the user-interface for setting shares/weights.
>  *      
>  * We increase resolution only if we have enough bits to allow this increased
>  * resolution (i.e. BITS_PER_LONG > 32). The costs for increasing resolution
>  * when BITS_PER_LONG <= 32 are pretty high and the returns do not justify the
>  * increased costs.
>  */
> #if 0 /* BITS_PER_LONG > 32 -- currently broken: it increases power usage under light load  */
> # define SCHED_LOAD_RESOLUTION  10
> # define scale_load(w)          ((w) << SCHED_LOAD_RESOLUTION)
> # define scale_load_down(w)     ((w) >> SCHED_LOAD_RESOLUTION)
>
> It mentioned some regressions, that's the history but
> sounds like a lot of testing is needed.

Thanks for the tip.  Yes, I was aware of this and bit worried about the
mentioned regression.  So I just wanted to increate the load weight of a
SCHED_IDLE task.

However I don't have a good test environment unfortunately so cannot
carry the work on at least for a while.

Thanks for your help anyway,
Namhyung