* [Eas-dev] Computing the dynamic-power-coefficient on Exynos5422
@ 2018-07-26 8:04 Oliver Effland
2018-07-27 2:06 ` Steven Miao
2018-08-02 16:14 ` Dietmar Eggemann
0 siblings, 2 replies; 6+ messages in thread
From: Oliver Effland @ 2018-07-26 8:04 UTC (permalink / raw)
To: linux-arm-kernel
Hi,
First of all: thanks for all the valuable input! @(Russell, Dietmar & Steven)
On 07/23/2018 08:51 AM, Steven Miao wrote:
> Hi Oliver,
>
> Generally these dynamic co-efficiency and static leakage power comes
> from hardware design team by running power estimation tool, however
> you can tun these co-efficiency and parameters for better performance.
>
> How did you get A7 or A15 power? Get Soc power or there's separate
> sensor for A7/A15?
Yes indeed, there are 2 seperate sensors for the A7 and A15.
> If you want to create cpu busy power model manually, you can:
> 1 leave 1 A7 + 1 A15 core online, keep other cores offline, disable
> thermal zone, disable idle state for all cores
> 2 set A7 frequency to 200Mhz, move all the task to A15, run cpu
> intensive task on A7(dhrystone, sysbench cpu test), read sensor to get
> 1 A7 power P1(read from A7 cluster sensor)
> 3 online 2 A7 core, run dhrystone on each core, and get power P2
>
> we can get 1 A7 core power at 200Mhz ~= P2 - P1 (A7 power X 2 + static
> power - A7 power X 1 - static power
I followed your manual steps (using a busy loop for 10s, instead of a benchmark) and corrected my calculations (with the static
power in mind):
A7:
freq(MHz) voltage(mV) Power(mW) A7_Power(mW) Stat_Power(mW) DPC
200 918.750 34.768 12.203 10.362 72.283955
400 920.000 60.805 21.25 18.305 62.765831
600 968.750 98.046 35.084 27.878 62.306652
800 1042.500 148.095 52.285 43.525 60.136063
1000 1116.250 212.214 76.112 59.99 61.084382
1200 1190.000 288.958 99.993 88.972 58.842948
1400 1288.750 411.936 144.807 122.322 62.276495
mean: 62
A15:
freq(MHz) voltage(mV) Power(mW) A15_Power(mW) Stat_Power(mW) DPC
200 916250 122.584 31.139 60.306 185.458477
500 916250 225.101 68.774 87.553 163.842401
800 942500 349.158 112.432 124.294 158.211202
1100 1016250 540.78 169.556 201.668 149.251725
1400 1078750 779.19 258.214 262.762 158.493016
1700 1180000 1179.344 392.107 395.13 165.650083
2000 1320000 1969.764 648.768 672.228 186.170798
mean: 164
The ratio between A15 and A7 is now 164/62 = 2.645
I hope the dynamic-power-coefficient's are more accurate now?
> dynamic-power-coefficient ~= 400 - 500 is reasonable for big core
> dynamic-power-coefficient ~= 100 is reasonable for LITTLE core.
> For IPA it can work effectively with a poor power model(not so
> accurate), for EAS it just need power model of LITTLE core and big
> core is proportional.
I'm guessing my values differ from your suggested values, because it's 32bit ARM?
Best regards,
Oliver
^ permalink raw reply [flat|nested] 6+ messages in thread
* [Eas-dev] Computing the dynamic-power-coefficient on Exynos5422
2018-07-26 8:04 [Eas-dev] Computing the dynamic-power-coefficient on Exynos5422 Oliver Effland
@ 2018-07-27 2:06 ` Steven Miao
2018-08-02 16:14 ` Dietmar Eggemann
1 sibling, 0 replies; 6+ messages in thread
From: Steven Miao @ 2018-07-27 2:06 UTC (permalink / raw)
To: linux-arm-kernel
Hi Oliver,
On Thu, Jul 26, 2018 at 4:05 PM Oliver Effland
<s5770874@msx.tu-dresden.de> wrote:
>
> Hi,
>
> First of all: thanks for all the valuable input! @(Russell, Dietmar & Steven)
>
> On 07/23/2018 08:51 AM, Steven Miao wrote:
> > Hi Oliver,
> >
> > Generally these dynamic co-efficiency and static leakage power comes
> > from hardware design team by running power estimation tool, however
> > you can tun these co-efficiency and parameters for better performance.
> >
> > How did you get A7 or A15 power? Get Soc power or there's separate
> > sensor for A7/A15?
> Yes indeed, there are 2 seperate sensors for the A7 and A15.
>
> > If you want to create cpu busy power model manually, you can:
> > 1 leave 1 A7 + 1 A15 core online, keep other cores offline, disable
> > thermal zone, disable idle state for all cores
> > 2 set A7 frequency to 200Mhz, move all the task to A15, run cpu
> > intensive task on A7(dhrystone, sysbench cpu test), read sensor to get
> > 1 A7 power P1(read from A7 cluster sensor)
> > 3 online 2 A7 core, run dhrystone on each core, and get power P2
> >
> > we can get 1 A7 core power at 200Mhz ~= P2 - P1 (A7 power X 2 + static
> > power - A7 power X 1 - static power
> I followed your manual steps (using a busy loop for 10s, instead of a benchmark) and corrected my calculations (with the static
> power in mind):
>
> A7:
> freq(MHz) voltage(mV) Power(mW) A7_Power(mW) Stat_Power(mW) DPC
> 200 918.750 34.768 12.203 10.362 72.283955
> 400 920.000 60.805 21.25 18.305 62.765831
> 600 968.750 98.046 35.084 27.878 62.306652
> 800 1042.500 148.095 52.285 43.525 60.136063
> 1000 1116.250 212.214 76.112 59.99 61.084382
> 1200 1190.000 288.958 99.993 88.972 58.842948
> 1400 1288.750 411.936 144.807 122.322 62.276495
>
> mean: 62
>
> A15:
> freq(MHz) voltage(mV) Power(mW) A15_Power(mW) Stat_Power(mW) DPC
> 200 916250 122.584 31.139 60.306 185.458477
> 500 916250 225.101 68.774 87.553 163.842401
> 800 942500 349.158 112.432 124.294 158.211202
> 1100 1016250 540.78 169.556 201.668 149.251725
> 1400 1078750 779.19 258.214 262.762 158.493016
> 1700 1180000 1179.344 392.107 395.13 165.650083
> 2000 1320000 1969.764 648.768 672.228 186.170798
>
> mean: 164
>
> The ratio between A15 and A7 is now 164/62 = 2.645
>
> I hope the dynamic-power-coefficient's are more accurate now?
>
> > dynamic-power-coefficient ~= 400 - 500 is reasonable for big core
> > dynamic-power-coefficient ~= 100 is reasonable for LITTLE core.
> > For IPA it can work effectively with a poor power model(not so
> > accurate), for EAS it just need power model of LITTLE core and big
> > core is proportional.
>
> I'm guessing my values differ from your suggested values, because it's 32bit ARM?
Yes, the ratio bit and LITTLE ~= 400/100 ~= 4.x is from A73 and A53 @
16nm process, it could vary from 32bit core or different process.
>
> Best regards,
> Oliver
Cheers,
Steven
^ permalink raw reply [flat|nested] 6+ messages in thread
* [Eas-dev] Computing the dynamic-power-coefficient on Exynos5422
2018-07-26 8:04 [Eas-dev] Computing the dynamic-power-coefficient on Exynos5422 Oliver Effland
2018-07-27 2:06 ` Steven Miao
@ 2018-08-02 16:14 ` Dietmar Eggemann
1 sibling, 0 replies; 6+ messages in thread
From: Dietmar Eggemann @ 2018-08-02 16:14 UTC (permalink / raw)
To: linux-arm-kernel
On 07/26/2018 10:04 AM, Oliver Effland wrote:
> Hi,
>
> First of all: thanks for all the valuable input! @(Russell, Dietmar & Steven)
>
> On 07/23/2018 08:51 AM, Steven Miao wrote:
[...]
> A7:
> freq(MHz) voltage(mV) Power(mW) A7_Power(mW) Stat_Power(mW) DPC
> 200 918.750 34.768 12.203 10.362 72.283955
> 400 920.000 60.805 21.25 18.305 62.765831
> 600 968.750 98.046 35.084 27.878 62.306652
> 800 1042.500 148.095 52.285 43.525 60.136063
> 1000 1116.250 212.214 76.112 59.99 61.084382
> 1200 1190.000 288.958 99.993 88.972 58.842948
> 1400 1288.750 411.936 144.807 122.322 62.276495
>
> mean: 62
>
> A15:
> freq(MHz) voltage(mV) Power(mW) A15_Power(mW) Stat_Power(mW) DPC
> 200 916250 122.584 31.139 60.306 185.458477
> 500 916250 225.101 68.774 87.553 163.842401
> 800 942500 349.158 112.432 124.294 158.211202
> 1100 1016250 540.78 169.556 201.668 149.251725
> 1400 1078750 779.19 258.214 262.762 158.493016
> 1700 1180000 1179.344 392.107 395.13 165.650083
> 2000 1320000 1969.764 648.768 672.228 186.170798
>
> mean: 164
>
> The ratio between A15 and A7 is now 164/62 = 2.645
>
> I hope the dynamic-power-coefficient's are more accurate now?
>
>> dynamic-power-coefficient ~= 400 - 500 is reasonable for big core
>> dynamic-power-coefficient ~= 100 is reasonable for LITTLE core.
>> For IPA it can work effectively with a poor power model(not so
>> accurate), for EAS it just need power model of LITTLE core and big
>> core is proportional.
>
> I'm guessing my values differ from your suggested values, because it's 32bit ARM?
We're running TC2 (32 bit Arm A15/A7 _test_ chip) with a ratio between
A15 and A7 of 990/113 = 7.44.
For EAS, i.e. for task wake-up placement decisions based on task
utilization and cpu capacity, on big.LITTLE this ratio is important.
How do you want to evaluate EAS v4 (or v5) on your ODROID-XU3 board?
What kind of workloads and/or benchmarks will you run on it?
-- Dietmar
[...]
^ permalink raw reply [flat|nested] 6+ messages in thread
* Computing the dynamic-power-coefficient on Exynos5422
@ 2018-07-20 14:15 Oliver Effland
2018-07-20 15:36 ` Russell King - ARM Linux
` (2 more replies)
0 siblings, 3 replies; 6+ messages in thread
From: Oliver Effland @ 2018-07-20 14:15 UTC (permalink / raw)
To: linux-arm-kernel
Hello everyone,
I hope this is the right place to ask, otherwise please just point me in the right direction.
I'm currently testing the EAS patches [v4] on an ODROID-XU3 board, which has an Exynos5422 SoC. However, the corresponding DT is missing the "dynamic-power-coefficient" that is needed for an appropriate EM.
So I'm trying to compute the dynamic-power-coefficient according to the formula:
Pdyn = dynamic-power-coefficient * V^2 * f
The frequency f is given by the DT.The actual Voltage and Power are determined by means of the on-chip sensors (returns the values for the specific cluster). When using the Voltage given in the DT, the differences for the d-p-coefficient are negligible.
So I'm calculating the d-p-coefficient (mW/MHz/uV^2) by reading out the following SoC sensor values.
For the little cluster (A7):
frequency(MHz) Voltage(V) Power(mW) Dynamic-power-coefficient
200 0.9175 49.470 ~2.938*10^-13
400 0.9165 91.892 ~2.736*10^-13
600 0.9638 149.454 ~2.682*10^-13
800 1.0263 223.453 ~2.652*10^-13
1000 1.1000 327.707 ~2.708*10^-13
1200 1.1725 445.899 ~2.703*10^-13
1400 1.2713 627.010 ~2.771*10^-13
For the big cluster (A15):
frequency(MHz) Voltage(V) Power(mW) Dynamic-power-coefficient
200 0.9162 159.676 ~9.510*10^-13
500 0.9138 325.480 ~7.797*10^-13
800 0.9288 511.360 ~7.410*10^-13
1100 1.0063 828.020 ~7.434*10^-13
1400 1.0713 1209.774 ~7.530*10^-13
1700 1.1750 1835.784 ~7.822*10^-13
2000 1.2700 2661.849 ~8.252*10^-13
But those values are way off for the DT, unless I multiply the d-p-coefficient with 10^15.
Assuming the Power value needs to be subtracted by a static component (power usage when idle), the results change as follows:
For the little cluster (A7):
frequency(MHz) Voltage(V) P-dyn(mW) Dynamic-power-coefficient
200 0.9175 36,381 ~2,160*10^-13
400 0.9165 73,438 ~2,186*10^-13
600 0.9638 122,470 ~2,197*10^-13
800 1.0263 184,135 ~2,185*10^-13
1000 1.1000 270,425 ~2,234*10^-13
1200 1.1725 367,121 ~2,225*10^-13
1400 1.2713 513,989 ~2,271*10^-13
For the big cluster (A15):
frequency(MHz) Voltage(V) P-dyn(mW) Dynamic-power-coefficient
200 0.9162 99,834 ~5,945*10^-13
500 0.9138 244,651 ~5,860*10^-13
800 0.9288 401,986 ~5,825*10^-13
1100 1.0063 664,565 ~5,966*10^-13
1400 1.0713 985,268 ~6,132*10^-13
1700 1.1750 1494,173 ~6,366*10^-13
2000 1.2700 2086,273 ~6,467*10^-13
While looking for examples in the kernel that calculate the dynamic-power-coefficient, I found this patch [1] by Caesar Wang that introduced this value for the rk3399 big cluster.
Unfortunately, I'm unable to reproduce the same results for the coefficient with the given values. My results are also with a 10^-13 factor and even when I scale them with 10^15, my results have a certain margin of error:
> frequency(MHz) Voltage(V) Current(mA) Dynamic-power-coefficient (My) d-p-coefficient
> 24 0.8 15
> 48 0.8 23 ~417 ~598
> 96 0.8 40 ~443 ~520
> 216 0.8 82 ~438 ~474
> 312 0.8 115 ~430 ~460
> 408 0.8 150 ~455 ~459
Is this the right approach, or am I missing something?
Hope someone can help me out.
Thanks in advance,
Oliver Effland
[1] https://patchwork.kernel.org/patch/9861505/
^ permalink raw reply [flat|nested] 6+ messages in thread* Computing the dynamic-power-coefficient on Exynos5422
2018-07-20 14:15 Oliver Effland
@ 2018-07-20 15:36 ` Russell King - ARM Linux
2018-07-30 16:32 ` [Eas-dev] " Vincent Guittot
2018-07-23 6:51 ` Steven Miao
2018-07-23 15:29 ` Dietmar Eggemann
2 siblings, 1 reply; 6+ messages in thread
From: Russell King - ARM Linux @ 2018-07-20 15:36 UTC (permalink / raw)
To: linux-arm-kernel
On Fri, Jul 20, 2018 at 04:15:30PM +0200, Oliver Effland wrote:
> Hello everyone,
>
> I hope this is the right place to ask, otherwise please just point me in the right direction.
>
> I'm currently testing the EAS patches [v4] on an ODROID-XU3 board, which has an Exynos5422 SoC. However, the corresponding DT is missing the "dynamic-power-coefficient" that is needed for an appropriate EM.
> So I'm trying to compute the dynamic-power-coefficient according to the formula:
>
> Pdyn = dynamic-power-coefficient * V^2 * f
>
> The frequency f is given by the DT.The actual Voltage and Power are determined by means of the on-chip sensors (returns the values for the specific cluster). When using the Voltage given in the DT, the differences for the d-p-coefficient are negligible.
>
> So I'm calculating the d-p-coefficient (mW/MHz/uV^2) by reading out the following SoC sensor values.
Well, let's work this forward. Let's take a simple, reasonable case.
V=1V, f = 200MHz. Now let's select the smallest value that
dynamic-power-coefficient can be in DT, which, as it's an integer,
is 1. We're told that V is in uV, and f is in MHz. So, let's plug
the figures in:
Pdyn = 1 * (1000000)^2 * 200
That is 2 * 10^14mW, or 200GW.
Somehow, I think the binding documentation is very wrong.
>
> For the little cluster (A7):
> frequency(MHz) Voltage(V) Power(mW) Dynamic-power-coefficient
> 200 0.9175 49.470 ~2.938*10^-13
> 400 0.9165 91.892 ~2.736*10^-13
> 600 0.9638 149.454 ~2.682*10^-13
> 800 1.0263 223.453 ~2.652*10^-13
> 1000 1.1000 327.707 ~2.708*10^-13
> 1200 1.1725 445.899 ~2.703*10^-13
> 1400 1.2713 627.010 ~2.771*10^-13
>
> For the big cluster (A15):
> frequency(MHz) Voltage(V) Power(mW) Dynamic-power-coefficient
> 200 0.9162 159.676 ~9.510*10^-13
> 500 0.9138 325.480 ~7.797*10^-13
> 800 0.9288 511.360 ~7.410*10^-13
> 1100 1.0063 828.020 ~7.434*10^-13
> 1400 1.0713 1209.774 ~7.530*10^-13
> 1700 1.1750 1835.784 ~7.822*10^-13
> 2000 1.2700 2661.849 ~8.252*10^-13
>
>
> But those values are way off for the DT, unless I multiply the d-p-coefficient with 10^15.
>
> Assuming the Power value needs to be subtracted by a static component (power usage when idle), the results change as follows:
>
> For the little cluster (A7):
> frequency(MHz) Voltage(V) P-dyn(mW) Dynamic-power-coefficient
> 200 0.9175 36,381 ~2,160*10^-13
> 400 0.9165 73,438 ~2,186*10^-13
> 600 0.9638 122,470 ~2,197*10^-13
> 800 1.0263 184,135 ~2,185*10^-13
> 1000 1.1000 270,425 ~2,234*10^-13
> 1200 1.1725 367,121 ~2,225*10^-13
> 1400 1.2713 513,989 ~2,271*10^-13
>
> For the big cluster (A15):
> frequency(MHz) Voltage(V) P-dyn(mW) Dynamic-power-coefficient
> 200 0.9162 99,834 ~5,945*10^-13
> 500 0.9138 244,651 ~5,860*10^-13
> 800 0.9288 401,986 ~5,825*10^-13
> 1100 1.0063 664,565 ~5,966*10^-13
> 1400 1.0713 985,268 ~6,132*10^-13
> 1700 1.1750 1494,173 ~6,366*10^-13
> 2000 1.2700 2086,273 ~6,467*10^-13
>
>
> While looking for examples in the kernel that calculate the dynamic-power-coefficient, I found this patch [1] by Caesar Wang that introduced this value for the rk3399 big cluster.
> Unfortunately, I'm unable to reproduce the same results for the coefficient with the given values. My results are also with a 10^-13 factor and even when I scale them with 10^15, my results have a certain margin of error:
>
> > frequency(MHz) Voltage(V) Current(mA) Dynamic-power-coefficient (My) d-p-coefficient
> > 24 0.8 15
> > 48 0.8 23 ~417 ~598
> > 96 0.8 40 ~443 ~520
> > 216 0.8 82 ~438 ~474
> > 312 0.8 115 ~430 ~460
> > 408 0.8 150 ~455 ~459
Hmm. Let's take your 48MHz values.
Pdyn (mW) = 0.8 (V) * 23 (mA) = 18.4mW
dynamic-power-coefficient = Pdyn / (V^2 * f)
= 18.4 (mW) / (800000 * 800000 * 48)
= 5.99 * 10^-13
Seems, again, to me that there is something very wrong with the
binding documentation.
Now, let's dig into the code:
* dev_pm_opp_get_voltage() - Gets the voltage corresponding to an opp
* Return: voltage in micro volt corresponding to the opp, else
* return 0
So this returns uV.
voltage_mv = dev_pm_opp_get_voltage(opp) / 1000;
This is mV.
u32 freq_mhz = freq_table[i].frequency / 1000;
MHz.
power = (u64)capacitance * freq_mhz * voltage_mv * voltage_mv;
do_div(power, 1000000000);
/* power is stored in mW */
freq_table[i].power = power;
So we have:
power(mW) = dynamic-power-coefficient * f(MHz) * V(mV)^2 / 10^9
That 10^9 can't come from the fact that the calculation is using mV
instead of uV. Let's see what happens if we use this:
dynamic-power-coefficient = power(mW) * 10^9 / (f(MHz) * V(mV)^2)
= 18.4 * 10^9 / (48 * 800 * 800)
= 599
which, given that dynamic-power-coefficient is an integer, is a damn
sight more sensible. This is equivalent to:
dynamic-power-coefficient = power(mW) * 10^15 / (f(MHz) * V(uV)^2)
which means the binding documentation is wrong.
Now, one of the things to bear in mind is that if you're measuring the
voltage and current, and wanting "dynamic power" then you don't get
that from multiplying the voltage and current - that includes the
_static_ consumption.
total power = static power + dynamic power
= V * (Ileakage + Idynamic)
and Idynamic will be based upon the frequency.
Taking your table:
> > frequency(MHz) Voltage(V) Current(mA) dpc Pdyn Ptot Pstatic
> > 24 0.8 15
> > 48 0.8 23 ~417 12.8 18.4 5.6
> > 96 0.8 40 ~443 27.2 32 4.8
> > 216 0.8 82 ~438 60.5 65.6 5.1
> > 312 0.8 115 ~430 85.9 92 6.1
> > 408 0.8 150 ~455 119 120 1
Pdyn comes from the dynamic-power-coefficient calculation, Ptot from
the voltage and current figures, and Pstatic being the difference
between Ptot and Pdyn.
Apart from the last, it looks like there's a static power of an average
5.4mW. Plugging that back in:
dynamic-power-coefficient = power(mW) * 10^9 / (f(MHz) * V(mV)^2)
= (18.4 - 5.4) * 10^9 / (48 * 800 * 800) = 423
= (32 - 5.4) * 10^9 / (96 * 800 * 800) = 433
= (65.6 - 5.4) * 10^9 / (216 * 800 * 800) = 435
= (92 - 5.4) * 10^9 / (312 * 800 * 800) = 434
= (120 - 5.4) * 10^9 / (408 * 800 * 800) = 439
which is a bit more believable (even compared to the original post
of the table.)
--
RMK's Patch system: http://www.armlinux.org.uk/developer/patches/
FTTC broadband for 0.8mile line in suburbia: sync at 13.8Mbps down 630kbps up
According to speedtest.net: 13Mbps down 490kbps up
^ permalink raw reply [flat|nested] 6+ messages in thread* [Eas-dev] Computing the dynamic-power-coefficient on Exynos5422
2018-07-20 15:36 ` Russell King - ARM Linux
@ 2018-07-30 16:32 ` Vincent Guittot
0 siblings, 0 replies; 6+ messages in thread
From: Vincent Guittot @ 2018-07-30 16:32 UTC (permalink / raw)
To: linux-arm-kernel
Hi Russel and Oliver,
On Fri, 20 Jul 2018 at 17:36, Russell King - ARM Linux
<linux@armlinux.org.uk> wrote:
>
> On Fri, Jul 20, 2018 at 04:15:30PM +0200, Oliver Effland wrote:
> > Hello everyone,
> >
> > I hope this is the right place to ask, otherwise please just point me in the right direction.
> >
> > I'm currently testing the EAS patches [v4] on an ODROID-XU3 board, which has an Exynos5422 SoC. However, the corresponding DT is missing the "dynamic-power-coefficient" that is needed for an appropriate EM.
> > So I'm trying to compute the dynamic-power-coefficient according to the formula:
> >
> > Pdyn = dynamic-power-coefficient * V^2 * f
> >
> > The frequency f is given by the DT.The actual Voltage and Power are determined by means of the on-chip sensors (returns the values for the specific cluster). When using the Voltage given in the DT, the differences for the d-p-coefficient are negligible.
> >
> > So I'm calculating the d-p-coefficient (mW/MHz/uV^2) by reading out the following SoC sensor values.
>
> Well, let's work this forward. Let's take a simple, reasonable case.
>
> V=1V, f = 200MHz. Now let's select the smallest value that
> dynamic-power-coefficient can be in DT, which, as it's an integer,
> is 1. We're told that V is in uV, and f is in MHz. So, let's plug
> the figures in:
>
> Pdyn = 1 * (1000000)^2 * 200
>
> That is 2 * 10^14mW, or 200GW.
>
> Somehow, I think the binding documentation is very wrong.
>
> >
> > For the little cluster (A7):
> > frequency(MHz) Voltage(V) Power(mW) Dynamic-power-coefficient
> > 200 0.9175 49.470 ~2.938*10^-13
> > 400 0.9165 91.892 ~2.736*10^-13
> > 600 0.9638 149.454 ~2.682*10^-13
> > 800 1.0263 223.453 ~2.652*10^-13
> > 1000 1.1000 327.707 ~2.708*10^-13
> > 1200 1.1725 445.899 ~2.703*10^-13
> > 1400 1.2713 627.010 ~2.771*10^-13
> >
> > For the big cluster (A15):
> > frequency(MHz) Voltage(V) Power(mW) Dynamic-power-coefficient
> > 200 0.9162 159.676 ~9.510*10^-13
> > 500 0.9138 325.480 ~7.797*10^-13
> > 800 0.9288 511.360 ~7.410*10^-13
> > 1100 1.0063 828.020 ~7.434*10^-13
> > 1400 1.0713 1209.774 ~7.530*10^-13
> > 1700 1.1750 1835.784 ~7.822*10^-13
> > 2000 1.2700 2661.849 ~8.252*10^-13
> >
> >
> > But those values are way off for the DT, unless I multiply the d-p-coefficient with 10^15.
> >
> > Assuming the Power value needs to be subtracted by a static component (power usage when idle), the results change as follows:
> >
> > For the little cluster (A7):
> > frequency(MHz) Voltage(V) P-dyn(mW) Dynamic-power-coefficient
> > 200 0.9175 36,381 ~2,160*10^-13
> > 400 0.9165 73,438 ~2,186*10^-13
> > 600 0.9638 122,470 ~2,197*10^-13
> > 800 1.0263 184,135 ~2,185*10^-13
> > 1000 1.1000 270,425 ~2,234*10^-13
> > 1200 1.1725 367,121 ~2,225*10^-13
> > 1400 1.2713 513,989 ~2,271*10^-13
> >
> > For the big cluster (A15):
> > frequency(MHz) Voltage(V) P-dyn(mW) Dynamic-power-coefficient
> > 200 0.9162 99,834 ~5,945*10^-13
> > 500 0.9138 244,651 ~5,860*10^-13
> > 800 0.9288 401,986 ~5,825*10^-13
> > 1100 1.0063 664,565 ~5,966*10^-13
> > 1400 1.0713 985,268 ~6,132*10^-13
> > 1700 1.1750 1494,173 ~6,366*10^-13
> > 2000 1.2700 2086,273 ~6,467*10^-13
> >
> >
> > While looking for examples in the kernel that calculate the dynamic-power-coefficient, I found this patch [1] by Caesar Wang that introduced this value for the rk3399 big cluster.
> > Unfortunately, I'm unable to reproduce the same results for the coefficient with the given values. My results are also with a 10^-13 factor and even when I scale them with 10^15, my results have a certain margin of error:
> >
> > > frequency(MHz) Voltage(V) Current(mA) Dynamic-power-coefficient (My) d-p-coefficient
> > > 24 0.8 15
> > > 48 0.8 23 ~417 ~598
> > > 96 0.8 40 ~443 ~520
> > > 216 0.8 82 ~438 ~474
> > > 312 0.8 115 ~430 ~460
> > > 408 0.8 150 ~455 ~459
>
> Hmm. Let's take your 48MHz values.
>
> Pdyn (mW) = 0.8 (V) * 23 (mA) = 18.4mW
>
> dynamic-power-coefficient = Pdyn / (V^2 * f)
> = 18.4 (mW) / (800000 * 800000 * 48)
> = 5.99 * 10^-13
>
> Seems, again, to me that there is something very wrong with the
> binding documentation.
>
> Now, let's dig into the code:
>
> * dev_pm_opp_get_voltage() - Gets the voltage corresponding to an opp
> * Return: voltage in micro volt corresponding to the opp, else
> * return 0
>
> So this returns uV.
>
> voltage_mv = dev_pm_opp_get_voltage(opp) / 1000;
>
> This is mV.
>
> u32 freq_mhz = freq_table[i].frequency / 1000;
>
> MHz.
>
> power = (u64)capacitance * freq_mhz * voltage_mv * voltage_mv;
> do_div(power, 1000000000);
>
> /* power is stored in mW */
> freq_table[i].power = power;
>
> So we have:
>
> power(mW) = dynamic-power-coefficient * f(MHz) * V(mV)^2 / 10^9
>
> That 10^9 can't come from the fact that the calculation is using mV
> instead of uV. Let's see what happens if we use this:
>
> dynamic-power-coefficient = power(mW) * 10^9 / (f(MHz) * V(mV)^2)
> = 18.4 * 10^9 / (48 * 800 * 800)
> = 599
>
> which, given that dynamic-power-coefficient is an integer, is a damn
> sight more sensible. This is equivalent to:
>
> dynamic-power-coefficient = power(mW) * 10^15 / (f(MHz) * V(uV)^2)
>
> which means the binding documentation is wrong.
I came to similar conclusion on another thread:
https://patchwork.kernel.org/patch/10493615
the right unit of dynamic-power-coefficient is mW/V^2/Ghz
Regards,
Vincent
>
> Now, one of the things to bear in mind is that if you're measuring the
> voltage and current, and wanting "dynamic power" then you don't get
> that from multiplying the voltage and current - that includes the
> _static_ consumption.
>
> total power = static power + dynamic power
> = V * (Ileakage + Idynamic)
>
> and Idynamic will be based upon the frequency.
>
> Taking your table:
>
> > > frequency(MHz) Voltage(V) Current(mA) dpc Pdyn Ptot Pstatic
> > > 24 0.8 15
> > > 48 0.8 23 ~417 12.8 18.4 5.6
> > > 96 0.8 40 ~443 27.2 32 4.8
> > > 216 0.8 82 ~438 60.5 65.6 5.1
> > > 312 0.8 115 ~430 85.9 92 6.1
> > > 408 0.8 150 ~455 119 120 1
>
> Pdyn comes from the dynamic-power-coefficient calculation, Ptot from
> the voltage and current figures, and Pstatic being the difference
> between Ptot and Pdyn.
>
> Apart from the last, it looks like there's a static power of an average
> 5.4mW. Plugging that back in:
>
> dynamic-power-coefficient = power(mW) * 10^9 / (f(MHz) * V(mV)^2)
> = (18.4 - 5.4) * 10^9 / (48 * 800 * 800) = 423
> = (32 - 5.4) * 10^9 / (96 * 800 * 800) = 433
> = (65.6 - 5.4) * 10^9 / (216 * 800 * 800) = 435
> = (92 - 5.4) * 10^9 / (312 * 800 * 800) = 434
> = (120 - 5.4) * 10^9 / (408 * 800 * 800) = 439
>
> which is a bit more believable (even compared to the original post
> of the table.)
>
> --
> RMK's Patch system: http://www.armlinux.org.uk/developer/patches/
> FTTC broadband for 0.8mile line in suburbia: sync at 13.8Mbps down 630kbps up
> According to speedtest.net: 13Mbps down 490kbps up
> _______________________________________________
> eas-dev mailing list
> eas-dev at lists.linaro.org
> https://lists.linaro.org/mailman/listinfo/eas-dev
^ permalink raw reply [flat|nested] 6+ messages in thread
* [Eas-dev] Computing the dynamic-power-coefficient on Exynos5422
2018-07-20 14:15 Oliver Effland
2018-07-20 15:36 ` Russell King - ARM Linux
@ 2018-07-23 6:51 ` Steven Miao
2018-07-23 15:29 ` Dietmar Eggemann
2 siblings, 0 replies; 6+ messages in thread
From: Steven Miao @ 2018-07-23 6:51 UTC (permalink / raw)
To: linux-arm-kernel
Hi Oliver,
Generally these dynamic co-efficiency and static leakage power comes
from hardware design team by running power estimation tool, however
you can tun these co-efficiency and parameters for better performance.
On Fri, Jul 20, 2018 at 10:16 PM Oliver Effland
<s5770874@msx.tu-dresden.de> wrote:
>
> Hello everyone,
>
> I hope this is the right place to ask, otherwise please just point me in the right direction.
>
> I'm currently testing the EAS patches [v4] on an ODROID-XU3 board, which has an Exynos5422 SoC. However, the corresponding DT is missing the "dynamic-power-coefficient" that is needed for an appropriate EM.
> So I'm trying to compute the dynamic-power-coefficient according to the formula:
>
> Pdyn = dynamic-power-coefficient * V^2 * f
>
> The frequency f is given by the DT.The actual Voltage and Power are determined by means of the on-chip sensors (returns the values for the specific cluster). When using the Voltage given in the DT, the differences for the d-p-coefficient are negligible.
>
> So I'm calculating the d-p-coefficient (mW/MHz/uV^2) by reading out the following SoC sensor values.
How did you get A7 or A15 power? Get Soc power or there's separate
sensor for A7/A15?
If you want to create cpu busy power model manually, you can:
1 leave 1 A7 + 1 A15 core online, keep other cores offline, disable
thermal zone, disable idle state for all cores
2 set A7 frequency to 200Mhz, move all the task to A15, run cpu
intensive task on A7(dhrystone, sysbench cpu test), read sensor to get
1 A7 power P1(read from A7 cluster sensor)
3 online 2 A7 core, run dhrystone on each core, and get power P2
we can get 1 A7 core power at 200Mhz ~= P2 - P1 (A7 power X 2 + static
power - A7 power X 1 - static power)
>
> For the little cluster (A7):
> frequency(MHz) Voltage(V) Power(mW) Dynamic-power-coefficient
> 200 0.9175 49.470 ~2.938*10^-13
> 400 0.9165 91.892 ~2.736*10^-13
> 600 0.9638 149.454 ~2.682*10^-13
> 800 1.0263 223.453 ~2.652*10^-13
> 1000 1.1000 327.707 ~2.708*10^-13
> 1200 1.1725 445.899 ~2.703*10^-13
> 1400 1.2713 627.010 ~2.771*10^-13
>
> For the big cluster (A15):
> frequency(MHz) Voltage(V) Power(mW) Dynamic-power-coefficient
> 200 0.9162 159.676 ~9.510*10^-13
> 500 0.9138 325.480 ~7.797*10^-13
> 800 0.9288 511.360 ~7.410*10^-13
> 1100 1.0063 828.020 ~7.434*10^-13
> 1400 1.0713 1209.774 ~7.530*10^-13
> 1700 1.1750 1835.784 ~7.822*10^-13
> 2000 1.2700 2661.849 ~8.252*10^-13
>
>
> But those values are way off for the DT, unless I multiply the d-p-coefficient with 10^15.
>
> Assuming the Power value needs to be subtracted by a static component (power usage when idle), the results change as follows:
>
> For the little cluster (A7):
> frequency(MHz) Voltage(V) P-dyn(mW) Dynamic-power-coefficient
> 200 0.9175 36,381 ~2,160*10^-13
> 400 0.9165 73,438 ~2,186*10^-13
> 600 0.9638 122,470 ~2,197*10^-13
> 800 1.0263 184,135 ~2,185*10^-13
> 1000 1.1000 270,425 ~2,234*10^-13
> 1200 1.1725 367,121 ~2,225*10^-13
> 1400 1.2713 513,989 ~2,271*10^-13
>
> For the big cluster (A15):
> frequency(MHz) Voltage(V) P-dyn(mW) Dynamic-power-coefficient
> 200 0.9162 99,834 ~5,945*10^-13
> 500 0.9138 244,651 ~5,860*10^-13
> 800 0.9288 401,986 ~5,825*10^-13
> 1100 1.0063 664,565 ~5,966*10^-13
> 1400 1.0713 985,268 ~6,132*10^-13
> 1700 1.1750 1494,173 ~6,366*10^-13
> 2000 1.2700 2086,273 ~6,467*10^-13
>
>
> While looking for examples in the kernel that calculate the dynamic-power-coefficient, I found this patch [1] by Caesar Wang that introduced this value for the rk3399 big cluster.
> Unfortunately, I'm unable to reproduce the same results for the coefficient with the given values. My results are also with a 10^-13 factor and even when I scale them with 10^15, my results have a certain margin of error:
>
> > frequency(MHz) Voltage(V) Current(mA) Dynamic-power-coefficient (My) d-p-coefficient
> > 24 0.8 15
> > 48 0.8 23 ~417 ~598
> > 96 0.8 40 ~443 ~520
> > 216 0.8 82 ~438 ~474
> > 312 0.8 115 ~430 ~460
> > 408 0.8 150 ~455 ~459
dynamic-power-coefficient ~= 400 - 500 is reasonable for big core
dynamic-power-coefficient ~= 100 is reasonable for LITTLE core.
For IPA it can work effectively with a poor power model(not so
accurate), for EAS it just need power model of LITTLE core and big
core is proportional.
>
>
> Is this the right approach, or am I missing something?
> Hope someone can help me out.
>
> Thanks in advance,
> Oliver Effland
>
> [1] https://patchwork.kernel.org/patch/9861505/
>
> _______________________________________________
> eas-dev mailing list
> eas-dev at lists.linaro.org
> https://lists.linaro.org/mailman/listinfo/eas-dev
-Steven
^ permalink raw reply [flat|nested] 6+ messages in thread
* [Eas-dev] Computing the dynamic-power-coefficient on Exynos5422
2018-07-20 14:15 Oliver Effland
2018-07-20 15:36 ` Russell King - ARM Linux
2018-07-23 6:51 ` Steven Miao
@ 2018-07-23 15:29 ` Dietmar Eggemann
2 siblings, 0 replies; 6+ messages in thread
From: Dietmar Eggemann @ 2018-07-23 15:29 UTC (permalink / raw)
To: linux-arm-kernel
Hi Oliver,
On 07/20/2018 04:15 PM, Oliver Effland wrote:
> Hello everyone,
>
> I hope this is the right place to ask, otherwise please just point me in the right direction.
>
> I'm currently testing the EAS patches [v4] on an ODROID-XU3 board, which has an Exynos5422 SoC. However, the corresponding DT is missing the "dynamic-power-coefficient" that is needed for an appropriate EM.
> So I'm trying to compute the dynamic-power-coefficient according to the formula:
>
> Pdyn = dynamic-power-coefficient * V^2 * f
>
> The frequency f is given by the DT.The actual Voltage and Power are determined by means of the on-chip sensors (returns the values for the specific cluster). When using the Voltage given in the DT, the differences for the d-p-coefficient are negligible.
>
> So I'm calculating the d-p-coefficient (mW/MHz/uV^2) by reading out the following SoC sensor values.
>
> For the little cluster (A7):
> frequency(MHz) Voltage(V) Power(mW) Dynamic-power-coefficient
> 200 0.9175 49.470 ~2.938*10^-13
> 400 0.9165 91.892 ~2.736*10^-13
> 600 0.9638 149.454 ~2.682*10^-13
> 800 1.0263 223.453 ~2.652*10^-13
> 1000 1.1000 327.707 ~2.708*10^-13
> 1200 1.1725 445.899 ~2.703*10^-13
> 1400 1.2713 627.010 ~2.771*10^-13
>
> For the big cluster (A15):
> frequency(MHz) Voltage(V) Power(mW) Dynamic-power-coefficient
> 200 0.9162 159.676 ~9.510*10^-13
> 500 0.9138 325.480 ~7.797*10^-13
> 800 0.9288 511.360 ~7.410*10^-13
> 1100 1.0063 828.020 ~7.434*10^-13
> 1400 1.0713 1209.774 ~7.530*10^-13
> 1700 1.1750 1835.784 ~7.822*10^-13
> 2000 1.2700 2661.849 ~8.252*10^-13
We do support Arm's TC2 platform (2xA15 and 3xA7) with the EAS project.
The latest EAS mainline integration includes Quentin's "[RFC PATCH v4
00/12] Energy Aware Scheduling" https://lkml.org/lkml/2018/6/28/301 .
You can find the appropriate patch for TC2 dynamic-power-coefficient
integration "arm: dts: vexpress-v2p-ca15_a7: Add
dynamic-power-coefficient properties" under
https://developer.arm.com/open-source/energy-aware-scheduling/eas-mainline-development
. Follow the 'eas/next/integration' link. The patch header has a little
bit of documentation about how we did it. We used the build-in TC2
Energy Meter.
This is all in sync with what Russel and Steve posted on this thread before.
Thanks for taking EAS for a spin on another Arm 32bit machine!
-- Dietmar
[...]
^ permalink raw reply [flat|nested] 6+ messages in thread
end of thread, other threads:[~2018-08-02 16:14 UTC | newest]
Thread overview: 6+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2018-07-26 8:04 [Eas-dev] Computing the dynamic-power-coefficient on Exynos5422 Oliver Effland
2018-07-27 2:06 ` Steven Miao
2018-08-02 16:14 ` Dietmar Eggemann
-- strict thread matches above, loose matches on Subject: below --
2018-07-20 14:15 Oliver Effland
2018-07-20 15:36 ` Russell King - ARM Linux
2018-07-30 16:32 ` [Eas-dev] " Vincent Guittot
2018-07-23 6:51 ` Steven Miao
2018-07-23 15:29 ` Dietmar Eggemann
This is an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.