The unit of dynamic-power-coefficient is described as mW/MHz/uV^2 whereas
its usage in the code assumes that unit is mW/GHz/V^2
In drivers/thermal/cpu_cooling.c, the code is :
power = (u64)capacitance * freq_mhz * voltage_mv * voltage_mv;
do_div(power, 1000000000);
which can be summarized as :
power (mW) = capacitance * freq_mhz/1000 * (voltage_mv/1000)^2
Furthermore, if we test basic values like :
voltage_mv = 1000mV = 1V
freq_mhz = 1000Mhz = 1Ghz
The minimum possible power, when dynamic-power-coefficient equals 1, will
be :
min power = 1 * 1000 * (1000000)^2 = 10^15 mW
which is not realistic
With the unit used by the code, the min power is
min power = 1 * 1 * 1^2 = 1mW which is far more realistic
Signed-off-by: Vincent Guittot <[email protected]>
---
Documentation/devicetree/bindings/arm/cpus.txt | 4 ++--
1 file changed, 2 insertions(+), 2 deletions(-)
diff --git a/Documentation/devicetree/bindings/arm/cpus.txt b/Documentation/devicetree/bindings/arm/cpus.txt
index 29e1dc5..0148d7d 100644
--- a/Documentation/devicetree/bindings/arm/cpus.txt
+++ b/Documentation/devicetree/bindings/arm/cpus.txt
@@ -274,7 +274,7 @@ described below.
Usage: optional
Value type: <prop-encoded-array>
Definition: A u32 value that represents the running time dynamic
- power coefficient in units of mW/MHz/uV^2. The
+ power coefficient in units of mW/GHz/V^2. The
coefficient can either be calculated from power
measurements or derived by analysis.
@@ -285,7 +285,7 @@ described below.
Pdyn = dynamic-power-coefficient * V^2 * f
- where voltage is in uV, frequency is in MHz.
+ where voltage is in V, frequency is in GHz.
Example 1 (dual-cluster big.LITTLE system 32-bit):
--
2.7.4
Hi Vincent,
Thanks for the patch. One comment about the choice of units below.
Vincent Guittot <[email protected]> writes:
> The unit of dynamic-power-coefficient is described as mW/MHz/uV^2 whereas
> its usage in the code assumes that unit is mW/GHz/V^2
Instead of choosing GHz as the base, I'd prefer to use uW/MHz/V^2. It'll
avoid introducing fractional GHz value for frequency calculations.
> In drivers/thermal/cpu_cooling.c, the code is :
>
> power = (u64)capacitance * freq_mhz * voltage_mv * voltage_mv;
> do_div(power, 1000000000);
>
> which can be summarized as :
> power (mW) = capacitance * freq_mhz/1000 * (voltage_mv/1000)^2
Which would then translate to -
power (mW) = power (uW) / 1000 = capacitance * freq_mhz * (voltage_mv/1000)^2
Thanks,
Punit
>
> Furthermore, if we test basic values like :
> voltage_mv = 1000mV = 1V
> freq_mhz = 1000Mhz = 1Ghz
>
> The minimum possible power, when dynamic-power-coefficient equals 1, will
> be :
> min power = 1 * 1000 * (1000000)^2 = 10^15 mW
> which is not realistic
>
> With the unit used by the code, the min power is
> min power = 1 * 1 * 1^2 = 1mW which is far more realistic
>
> Signed-off-by: Vincent Guittot <[email protected]>
> ---
> Documentation/devicetree/bindings/arm/cpus.txt | 4 ++--
> 1 file changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/Documentation/devicetree/bindings/arm/cpus.txt b/Documentation/devicetree/bindings/arm/cpus.txt
> index 29e1dc5..0148d7d 100644
> --- a/Documentation/devicetree/bindings/arm/cpus.txt
> +++ b/Documentation/devicetree/bindings/arm/cpus.txt
> @@ -274,7 +274,7 @@ described below.
> Usage: optional
> Value type: <prop-encoded-array>
> Definition: A u32 value that represents the running time dynamic
> - power coefficient in units of mW/MHz/uV^2. The
> + power coefficient in units of mW/GHz/V^2. The
> coefficient can either be calculated from power
> measurements or derived by analysis.
>
> @@ -285,7 +285,7 @@ described below.
>
> Pdyn = dynamic-power-coefficient * V^2 * f
>
> - where voltage is in uV, frequency is in MHz.
> + where voltage is in V, frequency is in GHz.
>
> Example 1 (dual-cluster big.LITTLE system 32-bit):
Hi Amit,
On Wed, 22 Aug 2018 at 12:11, Punit Agrawal <[email protected]> wrote:
>
> Hi Vincent,
>
> Thanks for the patch. One comment about the choice of units below.
>
> Vincent Guittot <[email protected]> writes:
>
> > The unit of dynamic-power-coefficient is described as mW/MHz/uV^2 whereas
> > its usage in the code assumes that unit is mW/GHz/V^2
>
> Instead of choosing GHz as the base, I'd prefer to use uW/MHz/V^2. It'll
> avoid introducing fractional GHz value for frequency calculations.
I don't understand your concern about fractional Ghz value for
frequency calculation ?
I mean, why it's a problem for frequency with Ghz vs Mhz but not a
problem for voltage with V vs mV ?
Don't we have the same "problem" in both case ?
>
> > In drivers/thermal/cpu_cooling.c, the code is :
> >
> > power = (u64)capacitance * freq_mhz * voltage_mv * voltage_mv;
> > do_div(power, 1000000000);
> >
> > which can be summarized as :
> > power (mW) = capacitance * freq_mhz/1000 * (voltage_mv/1000)^2
>
> Which would then translate to -
>
> power (mW) = power (uW) / 1000 = capacitance * freq_mhz * (voltage_mv/1000)^2
Not sure that the equation above is correct. If we consider uW/MHz/V^2
for the unit, the equation becomes :
power (mW) = power (uW) / 1000 = capacitance * freq_mhz *
(voltage_mv/1000)^2 / 1000
which can be rearranged as
power (mW) = power (uW) / 1000 = capacitance * freq_mhz/ 1000 *
(voltage_mv/1000)^2
TBH, I don't really mind between mW/GHz/V^2 or uW/MHz/V^2 as they are
the same at the end
but I don't catch your reasoning
Regards,
Vincent
>
> Thanks,
> Punit
>
> >
> > Furthermore, if we test basic values like :
> > voltage_mv = 1000mV = 1V
> > freq_mhz = 1000Mhz = 1Ghz
> >
> > The minimum possible power, when dynamic-power-coefficient equals 1, will
> > be :
> > min power = 1 * 1000 * (1000000)^2 = 10^15 mW
> > which is not realistic
> >
> > With the unit used by the code, the min power is
> > min power = 1 * 1 * 1^2 = 1mW which is far more realistic
> >
> > Signed-off-by: Vincent Guittot <[email protected]>
> > ---
> > Documentation/devicetree/bindings/arm/cpus.txt | 4 ++--
> > 1 file changed, 2 insertions(+), 2 deletions(-)
> >
> > diff --git a/Documentation/devicetree/bindings/arm/cpus.txt b/Documentation/devicetree/bindings/arm/cpus.txt
> > index 29e1dc5..0148d7d 100644
> > --- a/Documentation/devicetree/bindings/arm/cpus.txt
> > +++ b/Documentation/devicetree/bindings/arm/cpus.txt
> > @@ -274,7 +274,7 @@ described below.
> > Usage: optional
> > Value type: <prop-encoded-array>
> > Definition: A u32 value that represents the running time dynamic
> > - power coefficient in units of mW/MHz/uV^2. The
> > + power coefficient in units of mW/GHz/V^2. The
> > coefficient can either be calculated from power
> > measurements or derived by analysis.
> >
> > @@ -285,7 +285,7 @@ described below.
> >
> > Pdyn = dynamic-power-coefficient * V^2 * f
> >
> > - where voltage is in uV, frequency is in MHz.
> > + where voltage is in V, frequency is in GHz.
> >
> > Example 1 (dual-cluster big.LITTLE system 32-bit):
Vincent Guittot <[email protected]> writes:
> Hi Amit,
>
> On Wed, 22 Aug 2018 at 12:11, Punit Agrawal <[email protected]> wrote:
>>
>> Hi Vincent,
>>
>> Thanks for the patch. One comment about the choice of units below.
>>
>> Vincent Guittot <[email protected]> writes:
>>
>> > The unit of dynamic-power-coefficient is described as mW/MHz/uV^2 whereas
>> > its usage in the code assumes that unit is mW/GHz/V^2
>>
>> Instead of choosing GHz as the base, I'd prefer to use uW/MHz/V^2. It'll
>> avoid introducing fractional GHz value for frequency calculations.
>
> I don't understand your concern about fractional Ghz value for
> frequency calculation ?
> I mean, why it's a problem for frequency with Ghz vs Mhz but not a
> problem for voltage with V vs mV ?
> Don't we have the same "problem" in both case ?
You're right. It's the same problem in both cases.
>>
>> > In drivers/thermal/cpu_cooling.c, the code is :
>> >
>> > power = (u64)capacitance * freq_mhz * voltage_mv * voltage_mv;
>> > do_div(power, 1000000000);
>> >
>> > which can be summarized as :
>> > power (mW) = capacitance * freq_mhz/1000 * (voltage_mv/1000)^2
>>
>> Which would then translate to -
>>
>> power (mW) = power (uW) / 1000 = capacitance * freq_mhz * (voltage_mv/1000)^2
>
> Not sure that the equation above is correct. If we consider uW/MHz/V^2
> for the unit, the equation becomes :
> power (mW) = power (uW) / 1000 = capacitance * freq_mhz *
> (voltage_mv/1000)^2 / 1000
Yes, I missed the "/ 1000" at the end.
> which can be rearranged as
> power (mW) = power (uW) / 1000 = capacitance * freq_mhz/ 1000 *
> (voltage_mv/1000)^2
>
> TBH, I don't really mind between mW/GHz/V^2 or uW/MHz/V^2 as they are
> the same at the end
> but I don't catch your reasoning
The problem I was thinking of doesn't hold as it's the same issue with
voltage.
One benefit to go with uW/MHz/V^2 might be the extra resolution that it
provides. I'd prefer to go with uW/MHz/V^2 if there's no compelling
reason to go with anything else.
[...]
>> >
>> > Furthermore, if we test basic values like :
>> > voltage_mv = 1000mV = 1V
>> > freq_mhz = 1000Mhz = 1Ghz
>> >
>> > The minimum possible power, when dynamic-power-coefficient equals 1, will
>> > be :
>> > min power = 1 * 1000 * (1000000)^2 = 10^15 mW
>> > which is not realistic
>> >
>> > With the unit used by the code, the min power is
>> > min power = 1 * 1 * 1^2 = 1mW which is far more realistic
>> >
>> > Signed-off-by: Vincent Guittot <[email protected]>
>> > ---
>> > Documentation/devicetree/bindings/arm/cpus.txt | 4 ++--
>> > 1 file changed, 2 insertions(+), 2 deletions(-)
>> >
>> > diff --git a/Documentation/devicetree/bindings/arm/cpus.txt b/Documentation/devicetree/bindings/arm/cpus.txt
>> > index 29e1dc5..0148d7d 100644
>> > --- a/Documentation/devicetree/bindings/arm/cpus.txt
>> > +++ b/Documentation/devicetree/bindings/arm/cpus.txt
>> > @@ -274,7 +274,7 @@ described below.
>> > Usage: optional
>> > Value type: <prop-encoded-array>
>> > Definition: A u32 value that represents the running time dynamic
>> > - power coefficient in units of mW/MHz/uV^2. The
>> > + power coefficient in units of mW/GHz/V^2. The
>> > coefficient can either be calculated from power
>> > measurements or derived by analysis.
>> >
>> > @@ -285,7 +285,7 @@ described below.
>> >
>> > Pdyn = dynamic-power-coefficient * V^2 * f
>> >
>> > - where voltage is in uV, frequency is in MHz.
>> > + where voltage is in V, frequency is in GHz.
>> >
>> > Example 1 (dual-cluster big.LITTLE system 32-bit):
On Tue, 28 Aug 2018 at 16:42, Punit Agrawal <[email protected]> wrote:
>
> Vincent Guittot <[email protected]> writes:
>
> > Hi Amit,
> >
> > On Wed, 22 Aug 2018 at 12:11, Punit Agrawal <[email protected]> wrote:
> >>
> >> Hi Vincent,
> >>
> >> Thanks for the patch. One comment about the choice of units below.
> >>
> >> Vincent Guittot <[email protected]> writes:
> >>
> >> > The unit of dynamic-power-coefficient is described as mW/MHz/uV^2 whereas
> >> > its usage in the code assumes that unit is mW/GHz/V^2
> >>
> >> Instead of choosing GHz as the base, I'd prefer to use uW/MHz/V^2. It'll
> >> avoid introducing fractional GHz value for frequency calculations.
> >
> > I don't understand your concern about fractional Ghz value for
> > frequency calculation ?
> > I mean, why it's a problem for frequency with Ghz vs Mhz but not a
> > problem for voltage with V vs mV ?
> > Don't we have the same "problem" in both case ?
>
> You're right. It's the same problem in both cases.
>
> >>
> >> > In drivers/thermal/cpu_cooling.c, the code is :
> >> >
> >> > power = (u64)capacitance * freq_mhz * voltage_mv * voltage_mv;
> >> > do_div(power, 1000000000);
> >> >
> >> > which can be summarized as :
> >> > power (mW) = capacitance * freq_mhz/1000 * (voltage_mv/1000)^2
> >>
> >> Which would then translate to -
> >>
> >> power (mW) = power (uW) / 1000 = capacitance * freq_mhz * (voltage_mv/1000)^2
> >
> > Not sure that the equation above is correct. If we consider uW/MHz/V^2
> > for the unit, the equation becomes :
> > power (mW) = power (uW) / 1000 = capacitance * freq_mhz *
> > (voltage_mv/1000)^2 / 1000
>
> Yes, I missed the "/ 1000" at the end.
>
> > which can be rearranged as
> > power (mW) = power (uW) / 1000 = capacitance * freq_mhz/ 1000 *
> > (voltage_mv/1000)^2
> >
> > TBH, I don't really mind between mW/GHz/V^2 or uW/MHz/V^2 as they are
> > the same at the end
> > but I don't catch your reasoning
>
> The problem I was thinking of doesn't hold as it's the same issue with
> voltage.
>
> One benefit to go with uW/MHz/V^2 might be the extra resolution that it
> provides. I'd prefer to go with uW/MHz/V^2 if there's no compelling
> reason to go with anything else.
Ok, I'm going to update the patch with uW/MHz/V^2
Thanks
>
>
> [...]
>
> >> >
> >> > Furthermore, if we test basic values like :
> >> > voltage_mv = 1000mV = 1V
> >> > freq_mhz = 1000Mhz = 1Ghz
> >> >
> >> > The minimum possible power, when dynamic-power-coefficient equals 1, will
> >> > be :
> >> > min power = 1 * 1000 * (1000000)^2 = 10^15 mW
> >> > which is not realistic
> >> >
> >> > With the unit used by the code, the min power is
> >> > min power = 1 * 1 * 1^2 = 1mW which is far more realistic
> >> >
> >> > Signed-off-by: Vincent Guittot <[email protected]>
> >> > ---
> >> > Documentation/devicetree/bindings/arm/cpus.txt | 4 ++--
> >> > 1 file changed, 2 insertions(+), 2 deletions(-)
> >> >
> >> > diff --git a/Documentation/devicetree/bindings/arm/cpus.txt b/Documentation/devicetree/bindings/arm/cpus.txt
> >> > index 29e1dc5..0148d7d 100644
> >> > --- a/Documentation/devicetree/bindings/arm/cpus.txt
> >> > +++ b/Documentation/devicetree/bindings/arm/cpus.txt
> >> > @@ -274,7 +274,7 @@ described below.
> >> > Usage: optional
> >> > Value type: <prop-encoded-array>
> >> > Definition: A u32 value that represents the running time dynamic
> >> > - power coefficient in units of mW/MHz/uV^2. The
> >> > + power coefficient in units of mW/GHz/V^2. The
> >> > coefficient can either be calculated from power
> >> > measurements or derived by analysis.
> >> >
> >> > @@ -285,7 +285,7 @@ described below.
> >> >
> >> > Pdyn = dynamic-power-coefficient * V^2 * f
> >> >
> >> > - where voltage is in uV, frequency is in MHz.
> >> > + where voltage is in V, frequency is in GHz.
> >> >
> >> > Example 1 (dual-cluster big.LITTLE system 32-bit):