Re: Missing recalculation of scheduler tunables in case of cpu hot add/remove

[Christian Ehrhardt <ehrhardt@linux.vnet.ibm.com>]

Pavel Machek wrote:
> Hi!
>
>   
>>>>>> Aside from that, we probably should put an upper limit in place, as I
>>>>>> guess large cpu count machines get silly large values
>>>>>>       
>>>>>>             
>>>>> I agree to that, but in the code is already an upper limit of 
>>>>> 200.000.000 - well we might discuss if that is too low/high.
>>>>>     
>>>>>           
>>>> Yeah, I think we should cap it around the 8-16 CPUs.
>>>>
>>>>   
>>>>         
>>> ok for me, driven by that finding I think I have to measure different 
>>> kind of scalings anyway, but as usually that takes some time :-/
>>> At least too time much for the discussion & solution of that bug I guess.
>>>
>>> The question for now is what we do on cpu hot add/remove?
>>> Would hooking somewhere in kernel/cpu.c be the right approach - I'm not 
>>> quite sure about my own suggestion yet :-).
>>>       
>> Something like the below might work I suppose, just needs a cleanup and
>> such.
>>     
>
> I see a rather fundamental problem: what if user wants to override
> those values, and wants them stay that way
Yep a fundamental problem, but fortunately solved already ;-)

See the series "[PATCH 0/3] fix rescaling of scheduler tunables v2" 
posted after this discussion.
That is exactly what patch #2 is about.
Giving users the choice to either set things constant (scaling=none) or 
dynamic (log or linear) as it is done boot time.

As I considered it a bug to miss the updates, the current patch 
initializes it with scaling=log to let runtime and boot behave the same way.
I could do an update to better keep interfaces which would initialize it 
with "scaling=none" to reflect by default the behavior of the current 
code that is missing scaling completely.
Comments welcome

-- 

Grüsse / regards, Christian Ehrhardt
IBM Linux Technology Center, Open Virtualization