Re: [PATCH 2/2] cpufreq: powernv: Ramp-down global pstate slower than local-pstate

[Akshay Adiga <akshay.adiga@linux.vnet.ibm.com>]

Hi Balbir,

On 04/14/2016 11:10 AM, Balbir Singh wrote:

> On 13/04/16 04:06, Akshay Adiga wrote:
>> This patch brings down global pstate at a slower rate than the local
>> pstate. As the frequency transition latency from pmin to pmax is
>> observed to be in few millisecond granurality. It takes a performance
>> penalty during sudden frequency rampup. Hence by holding global pstates
>> higher than local pstate makes the subsequent rampups faster.
> What domains does local and global refer to?

The *global pstate* is a Chip-level entity as such, so the global entity
(Voltage)  is managed across several cores. But with a DCM (Dual Chip Modules),
its more of a socket-level entity and hence Voltage is managed across both chips.

The *local pstate* is a Core-level entity, so the local entity (frequency) is
managed across threads.

I will include this in the commit message. Thanks.

>
>> +/*
>> + * Quadratic equation which gives the percentage rampdown for time elapsed in
>> + * milliseconds. time can be between 0 and MAX_RAMP_DOWN_TIME ( milliseconds )
>> + * This equation approximates to y = -4e-6 x^2
> Thanks for documenting this, but I think it will also be good to explain why we
> use y = -4 e-6*x^2 as opposed to any other magic numbers.

Well it empirically worked out best this way.

On an idle system we want the Global Pstate to ramp-down from max value to min over
a span of ~5 secs. Also we want initially ramp-down slowly and ramp-down rapidly later
on, hence the equation.

I will try to make this in the comment more informative.

Regards

Akshay Adiga