enhanced intel speedstep feature was Re: speedstep-centrino on dothan

enhanced intel speedstep feature was Re: speedstep-centrino on dothan

[st3]

Pedro Ramalhais found another interesting thing in:
ftp://download.intel.com/design/Pentium4/manuals/25366816.pdf

that's the IA32_CLOCK_MODULATION (still called
MSR_IA32_THERM_CONTROL in include/asm-i386/msr.h, and I think this would
need a fix) register, that let set a reduced CPU duty cycle. So just today I
coded something to use it.

Just add to speedstep-centrino.c:
[yes I know, this isn't the Right Coding Style, but it's just to give an
idea of the whole thing]

static int sccm_level = 0;
module_param(sccm_level, uint, 0644);
MODULE_PARM_DESC(sccm_level, "Software Controlled Clock Modulation level: 0
	for disabling, 7 for 12.5\% duty cycle");

in the first part, right after the third #define

and this function after centrino_target():

static int centrino_target_sccm ()
{
        unsigned int msr, h;
        
        rdmsr(MSR_IA32_THERM_CONTROL, msr, h);
        msr &= (0<<1); msr &= (0<<2); msr &= (0<<3);

        switch (sccm_level) {
		case 0:
			msr &= (0<<4);
                case 1:
                        msr |= (1<<4); msr |= (1<<3); msr |= (1<<2);
			msr |= (1<<1); break;
		case 2:
                        msr |= (i<<4); msr |= (1<<3); msr |= (1<<2); break;
                case 3:
                        msr |= (1<<4); msr |= (1<<3); msr |= (1<<1); break;
                case 4:
                        msr |= (1<<4); msr |= (1<<3); break;
                case 5:
                        msr |= (1<<4); msr |= (1<<2); msr |= (1<<1); break;
                case 6:
                        msr |= (1<<4); msr |= (1<<2); break;
                case 7:
                        msr |= (1<<4); msr |= (1<<1); break;
        }

        wrmsr (MSR_IA32_THERM_CONTROL, msr, h);
}

Then you'll just have to insmod speedstep-centrino.ko sccm_level=x, where x
can be 0 to disable software cycle control modulation, 1 to enable a
reduced duty cycle of 87.5%, 7 to enable a duty cycle of 12.5% and so on.

Enabling, say, a duty cycle of 12.5% means that the CPU chip will be driven
by clock just one time every eight, thus reducing power consumption and
temperature (and it speeds down dramatically the CPU, too =).

I tested it so far on my Pentium M Dothan 715 and on a Dothan 725.

What do you think, could this be useful? I could provide a patch or a
stand-alone module. Should I write an interface to sysfs?


--
ciao
st3

Re: enhanced intel speedstep feature was Re: speedstep-centrino on dothan

[Dave Jones]

On Thu, Jul 07, 2005 at 10:00:27PM +0200, st3@riseup.net wrote:
  > Enabling, say, a duty cycle of 12.5% means that the CPU chip will be driven
 > by clock just one time every eight, thus reducing power consumption and
 > temperature (and it speeds down dramatically the CPU, too =).
 > 
 > I tested it so far on my Pentium M Dothan 715 and on a Dothan 725.
 > 
 > What do you think, could this be useful? I could provide a patch or a
 > stand-alone module. Should I write an interface to sysfs?

You just reinvented the p4-clockmod driver :-)
This hasn't been seen to save any power whatsoever that I've seen.
I've heard a few reports that it reduces heat for a few folks, but
betting on it increasing your battery life is probably a push.

		Dave

Re: enhanced intel speedstep feature was Re: speedstep-centrino on dothan

[st3]

On Thu, 7 Jul 2005 16:06:48 -0400
Dave Jones <davej@redhat.com> wrote:

> On Thu, Jul 07, 2005 at 10:00:27PM +0200, st3@riseup.net wrote:
>   > Enabling, say, a duty cycle of 12.5% means that the CPU chip will be driven
>  > by clock just one time every eight, thus reducing power consumption and
>  > temperature (and it speeds down dramatically the CPU, too =).
>  > 
>  > I tested it so far on my Pentium M Dothan 715 and on a Dothan 725.
>  > 
>  > What do you think, could this be useful? I could provide a patch or a
>  > stand-alone module. Should I write an interface to sysfs?
> 
> You just reinvented the p4-clockmod driver :-)

Oops. =)

> This hasn't been seen to save any power whatsoever that I've seen.

It drops down power rating by 1500-1800mW on my Toshiba Satellite A50
while idling at 400MHz.

> I've heard a few reports that it reduces heat for a few folks, but
> betting on it increasing your battery life is probably a push.

Just a latest question: can be p4-clockmod used together with
speedstep-centrino? If not, would it make any sense to patch
speedstep-centrino to use this feature too?


--
ciao
st3

Re: enhanced intel speedstep feature was Re: speedstep-centrino on dothan

[Dave Jones]

On Thu, Jul 07, 2005 at 10:22:25PM +0200, st3@riseup.net wrote:

 > > This hasn't been seen to save any power whatsoever that I've seen.
 > 
 > It drops down power rating by 1500-1800mW on my Toshiba Satellite A50
 > while idling at 400MHz.
 > 
 > > I've heard a few reports that it reduces heat for a few folks, but
 > > betting on it increasing your battery life is probably a push.
 > 
 > Just a latest question: can be p4-clockmod used together with
 > speedstep-centrino? If not, would it make any sense to patch
 > speedstep-centrino to use this feature too?

There was some discussion on cpufreq list long ago to make cpufreq
drivers stackable, but it's never been a high enough priority for
anyone to actually do the work.

		Dave

Re: enhanced intel speedstep feature was Re: speedstep-centrino on dothan

[Joseph Pingenot]

>Just a latest question: can be p4-clockmod used together with
>speedstep-centrino? If not, would it make any sense to patch
>speedstep-centrino to use this feature too?

I'm a little confused.  How is this different from the ACPI CPU throttling
  states (/proc/acpi/processor/CPUn/limit to set, throttling to see all
  T-states available)?
On my 1.5-year-old Pentium-M, frequency scaling and T-states are different
  beasties, and act entirely differently.  I'm currently in the process of
  rewriting my governor's brain to deal with the two more intelligently.

-Joseph

-- 
Joseph===============================================trelane@digitasaru.net
      Graduate Student in Physics, Freelance Free Software Developer

Re: enhanced intel speedstep feature was Re: speedstep-centrino on dothan

[Dominik Brodowski]

On Thu, Jul 07, 2005 at 03:51:17PM -0500, Joseph Pingenot wrote:
> >Just a latest question: can be p4-clockmod used together with
> >speedstep-centrino? If not, would it make any sense to patch
> >speedstep-centrino to use this feature too?
> 
> I'm a little confused.  How is this different from the ACPI CPU throttling
>   states (/proc/acpi/processor/CPUn/limit to set, throttling to see all
>   T-states available)?

T-states _tend_ to be utilized using chipset logic, while p4-clockmod is
done in-CPU.

> On my 1.5-year-old Pentium-M, frequency scaling and T-states are different
>   beasties, and act entirely differently.  I'm currently in the process of
>   rewriting my governor's brain to deal with the two more intelligently.

In your case, I would care about throttling. In very most cases it actually
increases energy consumption, as the state being entered is technically the
same to ACPI C2 (IIRC), so it is only "forced" idling and only useful if
"forced" idling is needed to not need active cooling.

	Dominik

Re: enhanced intel speedstep feature was Re: speedstep-centrino on dothan

[Joseph Pingenot]

From Dominik Brodowski on Thursday, 07 July, 2005:
>On Thu, Jul 07, 2005 at 03:51:17PM -0500, Joseph Pingenot wrote:
>> >Just a latest question: can be p4-clockmod used together with
>> >speedstep-centrino? If not, would it make any sense to patch
>> >speedstep-centrino to use this feature too?
>> I'm a little confused.  How is this different from the ACPI CPU throttling
>>   states (/proc/acpi/processor/CPUn/limit to set, throttling to see all
>>   T-states available)?
>T-states _tend_ to be utilized using chipset logic, while p4-clockmod is
>done in-CPU.
>> On my 1.5-year-old Pentium-M, frequency scaling and T-states are different
>>   beasties, and act entirely differently.  I'm currently in the process of
>>   rewriting my governor's brain to deal with the two more intelligently.
>In your case, I would care about throttling. In very most cases it actually
>increases energy consumption, as the state being entered is technically the
>same to ACPI C2 (IIRC), so it is only "forced" idling and only useful if
>"forced" idling is needed to not need active cooling.

Why would this cause more energy consumption?

-Joseph

-- 
trelane@digitasaru.net--------------------------------------------------
          Graduate student in physics, Free Software developer.

Re: enhanced intel speedstep feature was Re: speedstep-centrino on dothan

[Dominik Brodowski]

On Thu, Jul 07, 2005 at 04:34:14PM -0500, Joseph Pingenot wrote:
> >From Dominik Brodowski on Thursday, 07 July, 2005:
> >On Thu, Jul 07, 2005 at 03:51:17PM -0500, Joseph Pingenot wrote:
> >> >Just a latest question: can be p4-clockmod used together with
> >> >speedstep-centrino? If not, would it make any sense to patch
> >> >speedstep-centrino to use this feature too?
> >> I'm a little confused.  How is this different from the ACPI CPU throttling
> >>   states (/proc/acpi/processor/CPUn/limit to set, throttling to see all
> >>   T-states available)?
> >T-states _tend_ to be utilized using chipset logic, while p4-clockmod is
> >done in-CPU.
> >> On my 1.5-year-old Pentium-M, frequency scaling and T-states are different
> >>   beasties, and act entirely differently.  I'm currently in the process of
> >>   rewriting my governor's brain to deal with the two more intelligently.
> >In your case, I would care about throttling. In very most cases it actually
> >increases energy consumption, as the state being entered is technically the
> >same to ACPI C2 (IIRC), so it is only "forced" idling and only useful if
> >"forced" idling is needed to not need active cooling.
> 
> Why would this cause more energy consumption?

Let's say a specific computing task needs 1s to run at 100% CPU load. The
CPU consumes 22 W at normal operation, and 7.3 W when in ACPI C2 state which
is technically equivalent to throttling. [note: these are _real_ values from
a real datasheet for a real CPU you see in common usage]

If you're at 0% CPU throttling rate, you need 	22 Ws for this computing task,
if you're at 25%				24 Ws,
	  at 50%				29 Ws,
      and at 75%			that is 44 Ws for this computing task.

So for any sepcific computing task the energy consumption increases largely.

But what if the system idle otherwise?
If the CPU is put into an idle state the other time (e.g. there is no
other computing task for the CPU to do in a four-second span), it depends on
the idle state being used: if it is C2, these four seconds mean 44Ws of
energy being used; if it is C3, the CPU only needs 5.1Ws, so at 0% CPU
throttling you get 37 Ws; if it is C4, the CPU only needs 0.55Ws, so you
only get 24 Ws which is much less than the 44Ws you have at 75% throttling.

Please note that the weak spot in this calculation is the idle state the
CPU is forced to when doing idling. My investigations lead to the assumption
that it is the "Stop Grant" state on CPUs manufactored by Intel, which is
most often described by the ACPI C2 state ("Stop-Grant state"), while C3 is
"Deep Sleep State" or "Sleep State", and C4 is "Deeper Sleep State".

	Dominik

Re: enhanced intel speedstep feature was Re: speedstep-centrino on dothan

[Dominik Brodowski]

On Thu, Jul 07, 2005 at 10:22:25PM +0200, st3@riseup.net wrote:
> > This hasn't been seen to save any power whatsoever that I've seen.
> 
> It drops down power rating by 1500-1800mW on my Toshiba Satellite A50
> while idling at 400MHz.

Do you use ACPI-based idling? If so, in which state is the CPU in (cat
/proc/acpi/processor/*/power ? I suspect that you do not use ACPI (else 
you wouldn't need the table-based approach) or that the ACPI-based idling is
broken on your notebook; as then the Linux idle handler  only makes use of 
"hlt" (IIRC), that is ACPI C1, while throttling "forces" ACPI C2 (again
 IIRC).

	Dominik

Re: enhanced intel speedstep feature was Re: speedstep-centrino on dothan

[st3]

On Thu, 7 Jul 2005 23:10:33 +0200
Dominik Brodowski <linux@dominikbrodowski.net> wrote:

> Do you use ACPI-based idling? If so, in which state is the CPU in (cat
> /proc/acpi/processor/*/power ? I suspect that you do not use ACPI (else 
> you wouldn't need the table-based approach) or that the ACPI-based idling is
> broken on your notebook; as then the Linux idle handler  only makes use of 
> "hlt" (IIRC), that is ACPI C1, while throttling "forces" ACPI C2 (again
>  IIRC).

For idling, I didn't mean 'real idling', but instead just 'doing nothing'
in ACPI C1 state, that's simply a CPU usage < 1%.

Sorry for being so lame =)


--
ciao
st3

Re: enhanced intel speedstep feature was Re: speedstep-centrino on dothan

[Dominik Brodowski]

On Thu, Jul 07, 2005 at 11:22:38PM +0200, st3@riseup.net wrote:
> On Thu, 7 Jul 2005 23:10:33 +0200
> Dominik Brodowski <linux@dominikbrodowski.net> wrote:
> 
> > Do you use ACPI-based idling? If so, in which state is the CPU in (cat
> > /proc/acpi/processor/*/power ? I suspect that you do not use ACPI (else 
> > you wouldn't need the table-based approach) or that the ACPI-based idling is
> > broken on your notebook; as then the Linux idle handler  only makes use of 
> > "hlt" (IIRC), that is ACPI C1, while throttling "forces" ACPI C2 (again
> >  IIRC).
> 
> For idling, I didn't mean 'real idling', but instead just 'doing nothing'
> in ACPI C1 state, that's simply a CPU usage < 1%.
> 
> Sorry for being so lame =)

That's exactly the "idling" I meant. So if it is only ACPI C1, then
throttling makes some sense. It makes more sense, though, to get ACPI C2, C3
and possibly C4 to work :)

	Dominik

Re: speedstep-centrino on dothan

[st3]

On Wed, 06 Jul 2005 15:58:53 -0400
Bill Davidsen <davidsen@tmr.com> wrote:

> st3@riseup.net wrote:

[snip]

> > Moreover, I checked on Pentium M 725 and Pentium M 715 that the lowest
> > frequency at which the CPU can be set safely is not the 600MHz given in
> > datasheets, but 400MHz instead, with VID#A, VID#B, VID#C and VID#D (see
> > datasheet for more details) set to 0.908V.
> > 
> > I can provide a patch, let me know.
> 
> Slower is better

[snip]

Here's a quick trick for getting CPU to lower frequencies than the ones
read from ACPI tables, while still keeping them available.

Just add these lines in centrino_cpu_init_acpi(), in speedstep-centrino.c,
just after (and outside) the first 'for' cycle like this:
for (i=0; i<p.state_count; i++) { ...

centrino_model[cpu]->op_points[p.state_count - 1].index = 0x040D;
centrino_model[cpu]->op_points[p.state_count - 1].frequency = 400000;
p.states[p.state_count - 1].core_frequency = 400;

This will enable 400MHz at 908mV, which I found to be useful on Dothan
CPUs. If you want to get values for other frequencies/voltages, just use
something like:

#include <stdio.h>

int main() {
unsigned int index, frequency, voltage

index = (((frequency)/100) << 8) | ((voltage - 700) / 16);
printf ("%u\n", index);
}

frequency is expressed in MHz, voltage in mV, index is the value for
centrino_model[cpu]->op_points[y].index <- remember to change this y if you
want to add multiple frequencies besides the ACPI ones; for example, if you
want 500MHz at 940mV, you could add:

centrino_model[cpu]->op_points[p.state_count - 2].index = 0x1295;
centrino_model[cpu]->op_points[p.state_count - 2].index = 500000;
p.states[p.state_count - 2].core_frequency = 500;

Sorry if I have been off-topic.


--
ciao
st3

Re: speedstep-centrino on dothan

[Dominik Brodowski]

On Thu, Jul 07, 2005 at 11:59:28PM +0200, st3@riseup.net wrote:
> read from ACPI tables, while still keeping them available.

You're only keeping some of them available, as you overwrite one such
setting. Alternatively you can increase p.state_count by one early enough.

> index = (((frequency)/100) << 8) | ((voltage - 700) / 16);
> printf ("%u\n", index);
	printf ("0x%x\n", index);
is better

> want 500MHz at 940mV, you could add:
> 
> centrino_model[cpu]->op_points[p.state_count - 2].index = 0x1295;
> centrino_model[cpu]->op_points[p.state_count - 2].index = 500000;
						.frequency
> p.states[p.state_count - 2].core_frequency = 500;


	Dominik

Re: speedstep-centrino on dothan

[st3]

On Fri, 8 Jul 2005 07:10:50 +0200
Dominik Brodowski <linux@dominikbrodowski.net> wrote:

> On Thu, Jul 07, 2005 at 11:59:28PM +0200, st3@riseup.net wrote:
> > read from ACPI tables, while still keeping them available.
> 
> You're only keeping some of them available, as you overwrite one such
> setting. Alternatively you can increase p.state_count by one early enough.

Yes, but I was just overwriting the latest ones, that are never used =).
Anyway, your solution is theoretically better.


--
ciao
st3

Re: speedstep-centrino on dothan

[Nikolay Pelov]

On Thu, 7 Jul 2005 st3@riseup.net wrote:

> 
> int main() {
> unsigned int index, frequency, voltage
> 
> index = (((frequency)/100) << 8) | ((voltage - 700) / 16);
> printf ("%u\n", index);
> }
> 
> frequency is expressed in MHz, voltage in mV, index is the value for
> centrino_model[cpu]->op_points[y].index 

Few days ago I discussed on the cpufreq mailing list that the above formula
is not the correct way to compute the index. Instead of frequency, one
should use a multiplier, i.e.:

  index = (multiplier << 8) | ((voltage - 700) / 16);

and the multiplier is computed as multiplier = frequency/fsb_speed.
On older systems fsb_speed = 100 MHz and that's why the original formula
was working. But for the Sonoma (Centrino 2) Pentium M processors
fsb_speed = 133 MHz. 

To give an example, take a CPU speed of 800MHz and a front side bus of
133MHz. Using the original formula we will use the value

  frequency/100 = 800/100 = 8

which is wrong. The correct value is:

  multiplier = frequency/fsb_speed = 800/133 = 6.

Nikolay