cpufreq: powernow-k8 frequency transitions question

cpufreq: powernow-k8 frequency transitions question

[Gunter Ohrner]

Hi!

I'm currently reading and playing with the powernow-k8 cpufreq driver
module. However so far I wasn't able to find out how frequency transitions
are dealt with that have to be made using intermediate steps, ie.
transitions which cannot be done in one single step.

(I read about this at 
http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/26094.PDF
p. 269f.)

I found safety guards which return an error if an invalid frequency
transition ("lo-lo transition") is requested. Is this the only support for
these transitions currently implemented or did I miss something?
Is there more advanced support simply missing, or is the current behaviour
sufficient, and if yes, why does this suffice?

How do the conservative/ondemand cpufreq governors cope with the kernel
advertising frequencies it cannot directly transition between?

I'd be really cool if someone could enlighten me, just some pointers at
powernow-k8.c functions I should take a closer look at might already
help. :-)

Greetings,

  Gunter

-- 
Due to hw failure, I'm thinking buying a new laptop.  I would like to 
know what is the worst laptop with ACPI atm?        -- Linux ACPI 
driver hacker Bruno Ducrot looking for a "good" laptop... ;)
*** PGP-Verschlüsselung bei eMails erwünscht :-) *** PGP: 0x1128F25F ***

cpufreq: powernow-k8 frequency transitions question

[Gunter Ohrner]

Hi!

I already asked this question on linux-kernel@vger.kernel.org but maybe this
list would be a more appropriate place...

I'm currently reading and playing with the powernow-k8 cpufreq driver
module. However so far I wasn't able to find out how frequency transitions
are dealt with that have to be made using intermediate steps, ie.
transitions which cannot be done in one single step.

(I read about this at 
http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/26094.PDF
p. 269f.)

I found safety guards which return an error if an invalid frequency
transition ("lo-lo transition") is requested. Is this the only support for
these transitions currently implemented or did I miss something?
Is there more advanced support simply missing, or is the current behaviour
sufficient, and if yes, why does this suffice?

How do the conservative/ondemand cpufreq governors cope with the kernel
advertising frequencies it cannot directly transition between?

I'd be really cool if someone could enlighten me, just some pointers at
powernow-k8.c functions I should take a closer look at might already
help. :-)

Greetings,

  Gunter

-- 
One of the universal rules of happiness is: always be wary of any 
helpful item that weighs less than its operating manual.        -- 
(Terry Pratchett, Jingo)
*** PGP-Verschlüsselung bei eMails erwünscht :-) *** PGP: 0x1128F25F ***

Re: cpufreq: powernow-k8 frequency transitions question

[Andi Kleen]

Gunter Ohrner <G.Ohrner@post.rwth-aachen.de> writes:

> I already asked this question on linux-kernel@vger.kernel.org but maybe this
> list would be a more appropriate place...

The proceedings of the Ottawa Linux Symposium 2004 (available online) 
have a detailed paper on the powernow-k8 driver.

> I'm currently reading and playing with the powernow-k8 cpufreq driver
> module. However so far I wasn't able to find out how frequency transitions
> are dealt with that have to be made using intermediate steps, ie.
> transitions which cannot be done in one single step.
> 
> (I read about this at 
> http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/26094.PDF
> p. 269f.)
> 
> I found safety guards which return an error if an invalid frequency
> transition ("lo-lo transition") is requested. Is this the only support for
> these transitions currently implemented or did I miss something?
> Is there more advanced support simply missing, or is the current behaviour
> sufficient, and if yes, why does this suffice?
> 
> How do the conservative/ondemand cpufreq governors cope with the kernel
> advertising frequencies it cannot directly transition between?

Step by step. The frequency/voltage transistion functions 
in powernow-k8.c go through them in loops.

-Andi

Re: cpufreq: powernow-k8 frequency transitions question

[Gunter Ohrner]

Andi Kleen wrote:
>> I already asked this question on linux-kernel@vger.kernel.org but maybe
>> this list would be a more appropriate place...
> The proceedings of the Ottawa Linux Symposium 2004 (available online)
> have a detailed paper on the powernow-k8 driver.

Ah, thanks. I found it, read it, and will soon compare the description to
the current implementation.
From what I saw I already figured out the most important parts myself, as
the driver is rather readable, but the paper contained some interesting
additional details and reasons for some things I still wondered about.

>> How do the conservative/ondemand cpufreq governors cope with the kernel
>> advertising frequencies it cannot directly transition between?
> Step by step. The frequency/voltage transistion functions
> in powernow-k8.c go through them in loops.

Oh, ok, I must have missed that part in the driver, I will look for it.

Thanks a lot! :-)

Greetings,

  Gunter

-- 
The students were staring at her in the manner of those who have heard 
of the species 'female' but have never expected to get this close to 
one.        -- (Terry Pratchett, Soul Music)
*** PGP-Verschlüsselung bei eMails erwünscht :-) *** PGP: 0x1128F25F ***

RE: cpufreq: powernow-k8 frequency transitions question

[Devriendt, Paul]

> I found safety guards which return an error if an invalid frequency
> transition ("lo-lo transition") is requested.

Transitions between low frequencies (<=1.4GHz) are not supported by
hardware. The way to do them is to transition to an intermediate
high frequency. No BIOSs define two low PStates so the case never
occurs, but it could theoretically be done. The AMD BIOS and Kernel
Developers Guide available on AMD's web site provides more information
on K8 PStates.

Paul.

RE: cpufreq: powernow-k8 frequency transitions question

[Gunter Ohrner]

Devriendt, Paul wrote:
> Transitions between low frequencies (<=1.4GHz) are not supported by
> hardware. The way to do them is to transition to an intermediate
> high frequency.

Yes, I read about that in the PDF I referred to. I guess that#s the guide
you also mentioned?

> No BIOSs define two low PStates so the case never occurs, but it could
> theoretically be done. 

I'm currently experimenting with the speeds my CPU can run at. (Out of spec,
at lower frequencies and voltages to safe a few watts and keep everything
cool'n'quiet, so to say... ;)

It seems to run fine a several low pstates, I just need to check if
supporting low-low-transitions would be worth the hassle, or if I just
limit my pstate table to two pstates between which the CPU can transition
directly.

Thanks a lot for your answer,

  Gunter

-- 
Thunder rolled. ... It rolled a six.        -- (Terry Pratchett, 
Guards! Guards!)
*** PGP-Verschlüsselung bei eMails erwünscht :-) *** PGP: 0x1128F25F ***

Re: cpufreq: powernow-k8 frequency transitions question

[Bruno Ducrot]

On Mon, Jan 02, 2006 at 11:37:12AM -0600, Devriendt, Paul wrote:
> > I found safety guards which return an error if an invalid frequency
> > transition ("lo-lo transition") is requested.
> 
> Transitions between low frequencies (<=1.4GHz) are not supported by
> hardware. The way to do them is to transition to an intermediate
> high frequency. No BIOSs define two low PStates so the case never
> occurs, but it could theoretically be done. The AMD BIOS and Kernel
> Developers Guide available on AMD's web site provides more information
> on K8 PStates.

I'm wondering if this would have some more power saving if there are more
low p-states.  I should some day look at this more seriously.

-- 
Bruno Ducrot

--  Which is worse:  ignorance or apathy?
--  Don't know.  Don't care.

Re: cpufreq: powernow-k8 frequency transitions question

[Gunter Ohrner]

Bruno Ducrot wrote:
> I'm wondering if this would have some more power saving if there are more
> low p-states.  I should some day look at this more seriously.

I own a Winchester 3000+ (max. 1,8GHz) which I tried to operate at 800 MHz
(instead of 1,0GHz) but it locked hard. It can be forced to intermediate
out-of-spec speeds (1,2GHz, 1,4GHz and 1,6GHz, although 1,4 should not be
reachable according to AMDs frequency transition spec).

Though I'm not sure if there's any win from these unsupported P-states as
1,0GHz should suffice for most tasks and for everything else the system can
quickly toggle to 1,8 GHz. I currently can't think of any task which would
require more processing power than provided by 1,0GHz but less than
provided by 1,8GHz over a longer period of time, so that the intermediate
P-states would actually be used by the ondemand governor for longer than a
fraction of a second... Though it could be interesting for faster CPUs than
mine that are decoding HD video or similar.

Greetings,

  Gunter

-- 
Ever noticed that people who spend money on beer, cigarettes and drugs 
are always complaining about being broke and not feeling well?        
-- Unbekannter Autor
*** PGP-Verschlüsselung bei eMails erwünscht :-) *** PGP: 0x1128F25F ***

Re: cpufreq: powernow-k8 frequency transitions question

[Nebojsa Trpkovic]

The most usefull powersaving is achieved by lowering the voltage and not 
the further lowering of frequency.
Power consumption (and heat dissipation) rise almost exactly with the 
square of core voltage, so one could save a lot of power by lowering 
voltage even just a little bit.
 I've found that almost all Athlon64 CPUs can work at a lot lower 
voltage then in AMD specs.

Newcastle 3000+   p-states: 1.0GHz @1.1V, 1.8GHz @1.4V, 2.0GHz @1.5V
http://www.aaen.edu.yu/~tnt/newcastle.png

Winchester 3000+   p-states: 1.0GHz @1.1V, 1.8GHz @1.4V
http://www.aaen.edu.yu/~tnt/winchester.png

So, I've modified powernow-k8.c to lower the voltages read from BIOS for 
every p-state. Although I've made graphs of stable voltages by running 
prime95 for at least 12 hours in every state,  I've left 0.1V over the 
minimum stable voltage just to be 110% sure it will work stable:
Winchester runs with 1.0GHz @1.0V and 1.8GHz @1.275V

Winchester is turned on 24/7 for last 6 months as it serves one LAN and 
one wifi network (DNS, Proxy, firewall, web, teamspeak2, gentoo rsync, 
ftp, 1TB samba...) and it has NO problems at all!

Newcastle was my desktop CPU and now there's Palermo Sempron 2800+.

Greetings,
Nebojsa Trpkovic



Gunter Ohrner wrote:

>Bruno Ducrot wrote:
>  
>
>>I'm wondering if this would have some more power saving if there are more
>>low p-states.  I should some day look at this more seriously.
>>    
>>
>
>I own a Winchester 3000+ (max. 1,8GHz) which I tried to operate at 800 MHz
>(instead of 1,0GHz) but it locked hard. It can be forced to intermediate
>out-of-spec speeds (1,2GHz, 1,4GHz and 1,6GHz, although 1,4 should not be
>reachable according to AMDs frequency transition spec).
>
>Though I'm not sure if there's any win from these unsupported P-states as
>1,0GHz should suffice for most tasks and for everything else the system can
>quickly toggle to 1,8 GHz. I currently can't think of any task which would
>require more processing power than provided by 1,0GHz but less than
>provided by 1,8GHz over a longer period of time, so that the intermediate
>P-states would actually be used by the ondemand governor for longer than a
>fraction of a second... Though it could be interesting for faster CPUs than
>mine that are decoding HD video or similar.
>
>Greetings,
>
>  Gunter
>
>  
>

Re: cpufreq: powernow-k8 frequency transitions question

[Bruno Ducrot]

On Wed, Jan 04, 2006 at 02:50:55AM +0100, Nebojsa Trpkovic wrote:
> The most usefull powersaving is achieved by lowering the voltage and not 
> the further lowering of frequency.

Of course.

> Power consumption (and heat dissipation) rise almost exactly with the 
> square of core voltage, so one could save a lot of power by lowering 
> voltage even just a little bit.

I think I can understand your point.

> I've found that almost all Athlon64 CPUs can work at a lot lower 
> voltage then in AMD specs.

You are right.  But I am not sure however if running out-of-spec a
processor will have a consequence for it's whole life time though.

> Newcastle 3000+   p-states: 1.0GHz @1.1V, 1.8GHz @1.4V, 2.0GHz @1.5V
> http://www.aaen.edu.yu/~tnt/newcastle.png
> 
> Winchester 3000+   p-states: 1.0GHz @1.1V, 1.8GHz @1.4V
> http://www.aaen.edu.yu/~tnt/winchester.png
> 
> So, I've modified powernow-k8.c to lower the voltages read from BIOS for 
> every p-state. Although I've made graphs of stable voltages by running 
> prime95 for at least 12 hours in every state,  I've left 0.1V over the 
> minimum stable voltage just to be 110% sure it will work stable:
> Winchester runs with 1.0GHz @1.0V and 1.8GHz @1.275V
> 
> Winchester is turned on 24/7 for last 6 months as it serves one LAN and 
> one wifi network (DNS, Proxy, firewall, web, teamspeak2, gentoo rsync, 
> ftp, 1TB samba...) and it has NO problems at all!
> 
> Newcastle was my desktop CPU and now there's Palermo Sempron 2800+.
> 

My though was more towards what Gunter Ohrner stated, that is if it's
interresting to add some intermediate low p-state between the already
given one and the first high p-state given the actual status of
the algorithms used (as the on-demand and maybe other I'm working ATM).

> 
> 
> Gunter Ohrner wrote:
> 
> >Bruno Ducrot wrote:
> > 
> >
> >>I'm wondering if this would have some more power saving if there are more
> >>low p-states.  I should some day look at this more seriously.
> >>   
> >>
> >
> >I own a Winchester 3000+ (max. 1,8GHz) which I tried to operate at 800 MHz
> >(instead of 1,0GHz) but it locked hard. It can be forced to intermediate
> >out-of-spec speeds (1,2GHz, 1,4GHz and 1,6GHz, although 1,4 should not be
> >reachable according to AMDs frequency transition spec).
> >
> >Though I'm not sure if there's any win from these unsupported P-states as
> >1,0GHz should suffice for most tasks and for everything else the system can
> >quickly toggle to 1,8 GHz. I currently can't think of any task which would
> >require more processing power than provided by 1,0GHz but less than
> >provided by 1,8GHz over a longer period of time, so that the intermediate
> >P-states would actually be used by the ondemand governor for longer than a
> >fraction of a second... Though it could be interesting for faster CPUs than
> >mine that are decoding HD video or similar.
> >
> >Greetings,
> >
> > Gunter
> >
> > 
> >
> 
> _______________________________________________
> Cpufreq mailing list
> Cpufreq@lists.linux.org.uk
> http://lists.linux.org.uk/mailman/listinfo/cpufreq

-- 
Bruno Ducrot

--  Which is worse:  ignorance or apathy?
--  Don't know.  Don't care.

Re: cpufreq: powernow-k8 frequency transitions question

[Nebojsa Trpkovic]

Bruno Ducrot wrote:

>My though was more towards what Gunter Ohrner stated, that is if it's
>interresting to add some intermediate low p-state between the already
>given one and the first high p-state given the actual status of
>the algorithms used (as the on-demand and maybe other I'm working ATM).
>
I just wanted to tell Gunter Ohrner that there are some more usefull 
methods then lowering CPU frequency to 800MHz.

As for low-low p-state transition, my experiments show that it's quite 
possible, but I don't know exact details such as transition latency and 
so on. Namely, when I've got Newcastle and Winchster, I've wanted to 
test them on various frequencies and voltages to make some kind of 
undetailed "preview" of graphs I made latter. It was a little bit boring 
to restart computer every time I wanted to change frequency and/or 
voltage, so I've installed Windows XP and software called "CPU MSR" wich 
allowed me to manually set FID and VID of the CPU in the real-time 
operation. (I think that FID transition latency was about 1000 "bus 
cycles" and VID transition latency about 1ms, but I'm not sure). So, on 
the Newcastle I've menaged to change multiplier down to 4 (800MHz), 
because my Newcastle was Socket 754 CPU and had 800MHz HTT. All other 
transitions between low p-states went without problem (multiplier 4 to 
5, 5 to 6, 4 to 6, 6 to 4 and so on).

>>I've found that almost all Athlon64 CPUs can work at a lot lower 
>>voltage then in AMD specs.
>>    
>>
>
>You are right.  But I am not sure however if running out-of-spec a
>processor will have a consequence for it's whole life time though.
>  
>
I can understand that running processor out-of-spec in terms of higher 
voltage than default one in long time periods could be dangerous because 
of heat disipation, changing of internal structure due to "runaway" 
atoms pulled away by electromagnetic forces stronger then at default 
voltage and so on,
but I realy can't see and negative aspect nor physical threat for 
processor when it's undervolted, except (ofcorse) stabilty issues if one 
goes too far with voltage lowering (where prime95 comes very handy :) ).

Greetings,
Nebojsa Trpkovic

Re: cpufreq: powernow-k8 frequency transitions question

[Gunter Ohrner]

Bruno Ducrot wrote:
> You are right.  But I am not sure however if running out-of-spec a
> processor will have a consequence for it's whole life time though.

Mh, I'm no processor expert, but how could underclocking and/or undervolting
cause any harm to the hardware? Overclocking and using higher voltages is
obvious, it increases power- and heat dissipation and increases aging of
the CPU's materials. However, does anyone here know if lowering the
frequencies and voltages could have any ill effects to the hardware?

Greetings,

  Gunter

-- 
"I used to think that *I* was stupid, and then I met philosophers."
        -- (Terry Pratchett, Small Gods)
*** PGP-Verschlüsselung bei eMails erwünscht :-) *** PGP: 0x1128F25F ***

Re: cpufreq: powernow-k8 frequency transitions question

[Gunter Ohrner]

Nebojsa Trpkovic wrote:
> The most usefull powersaving is achieved by lowering the voltage and not 
> the further lowering of frequency.

Yes, but lower clock speeds usually allow for lower voltages, so that's what
you gain by lowering the core frequency.

> minimum stable voltage just to be 110% sure it will work stable:
> Winchester runs with 1.0GHz @1.0V and 1.8GHz @1.275V

I'm currently running my Winchester 3000+ at 1,0GHZ@1,0V and 1,8GHz@1,15V
only, and it's rock solid for me so far - it survived hours of 100% cpu
load during normal operation and using cpuburn's burnK7, either locked at
on of the frequencies or in dynamic mode using the ondemand governor.

> So, I've modified powernow-k8.c to lower the voltages read from BIOS for 
> every p-state.

Changing the vlaues returned by the BIOS by a function is a really
interesting approach, especially if there's not much point in providing
additional P-states, which is my current impression. At the moment I have
ACPI table queries turned off in powernow-k8 and supply a hand craftes PSB
table, but obviously that's not the most flexible solution. Mh, two VID and
FID mapping function's which initially are identity functions but with some
sysfs-tuneable parameters sounds interesting to me... :-)

I guess I should try that.

Greetings and thanks for that idea,

  Gunter

-- 
The question seldom addressed is *where* Medusa had snakes. Underarm 
hair is an even more embarassing problem when it keeps biting the top 
of the deodorant bottle.        -- (Terry Pratchett, Soul Music)
*** PGP-Verschlüsselung bei eMails erwünscht :-) *** PGP: 0x1128F25F ***

Re: cpufreq: powernow-k8 frequency transitions question

[Gunter Ohrner]

Bruno Ducrot wrote:
> I'm wondering if this would have some more power saving if there are more
> low p-states.  I should some day look at this more seriously.

Ok, I just took some time to view a live concert DVD instead of studying or
coding... ;)

Well, as I result of that I changed my mind and must say: it might. At least
for saving a bit power during video decoding. I currently cannot imagine
different processing needs which require more processing power than
available at 1,0GHz and less than maximum, but there might be.

Video decoding at least needs much more processor cycles than I anticipated,
altought this DVD probably was pretty much the worst-case DVD one can
currently think of - rapid cuts, flashing lights, much movement and
interlaced video... :-/

Including deinterlacing and q2 or q3 pp filter postprocessing xine used
about 70% processing time at 1,0GHz averaged over a second (I didn't do
precise logging but just watched gkrellm with half an eye... ;); but very
"chunky" with lots of short peaks. Obviously xine doesn't use the
opportunity of available processing power to decode frames in advance and
even that out a bit, at least that was my impression.

There where some scenes where ondemand actually switched the CPU to 1,8GHz
and there where quite a few a fraction-of-a-second-peaks which seemed to be
too short for a speed transition but where the decoder would obviously have
needed a bit additional performance.

And I imagine that if even a DVD sometimes needs a bit more performance than
1,0GHz at my CPU can deliver, high-res xvid like videos will do as well,
probably even worse. 1,8 GHz would still be too much, so using faster low
states sounds reasonable for these tasks.

I did some basic energy consumption measurements when playing around with my
clock speeds last week, I just attached my whole computer to an el-cheapo
wattmeter.

Running at standard voltages the system running at 1,0GHz@1,1V needed 5
watts less compared to 1,8GHz@1,4V if idle and about 25 watts less if
loaded.

However, running at my lower voltages there where (only) 2 watts difference
when idle and about 12 watts under load.

In both cases I used burnK7 to load the system as it caused higher power
consumption than piping /dev/zero through gzip to /dev/null.

Now the question is if anyone considers these differences sufficiently
significant to invest time into tweaking the pnk8 driver... :-)

Greetings,

  Gunter

-- 
A nerd is someone whose life revolves around computers and technology. 
A geek is someone whose life revolves around computers and technology, 
and likes it!!
*** PGP-Verschlüsselung bei eMails erwünscht :-) *** PGP: 0x1128F25F ***

RE: cpufreq: powernow-k8 frequency transitions question

[Devriendt, Paul]

> I own a Winchester 3000+ (max. 1,8GHz) which I tried to 
> operate at 800 MHz
> (instead of 1,0GHz) but it locked hard.

Core speed must be >= HyperTransport link speed. It is possible
that you have a HT link running at 1 GHz.

Paul.