[Xenomai-help] Frequency downscaling of new intel CPU

[Xenomai-help] Frequency downscaling of new intel CPU

[Petr Cervenka]

Hello.

I recently tried newer computer and I suprusingly realized that it is bigger load than the old one with our xenomai application.
The problem is in the frequency downscaling of the processor when the load in under 100%.
Of course I have disabled CPU_FREQ, ACPI_PROCESSOR and INTEL_IDLE as suggested.
Later I tried also to disable CPU_IDLE, 7300_IDLE (intel chipset idle memory) and all power efficiency settings in the BIOS, but without success.
The experienced behaviour is following:
when the process is waiting for an event, the processor lowers its frequency and when the event happens it tries to restore (slowly) its nominal frequency.

Do you have any advice or tip what to try, because I'm really desperate?

Best regards

Petr Cervenka

configuration:
Xeon E3-1220 (Sandy Bridge)
linux-2.6.38.8
xenomai-2.5.6

Re: [Xenomai-help] Frequency downscaling of new intel CPU

[Petr Cervenka]

Hello again.
I know, that my problem is not so close related with the xenomai. But I think many of xenomai users with new intel CPUs have similar problem.
I tried it it with two different main boards and none of the BIOS settings helped. (C-states: disabled, Speedstep: disabled, CPU idle: high performance, TurboBoost: disabled, ...). My theory is that the linux kernel overrides BIOS settings, but I don't know how to prove it.
Any ideas that could help me?

Petr Cervenka

______________________________________________________________.
> Od: "Petr Cervenka" <grugh@domain.hid>
> Komu: "xenomai-help" <xenomai@xenomai.org>
> Datum: 26.09.2011 09:02
> Předmět: Frequency downscaling of new intel CPU
>
>Hello.
>
>I recently tried newer computer and I suprusingly realized that it has bigger load than the old one with our xenomai application.
>The problem is in the frequency downscaling of the processor when the load in under 100%.
>Of course I have disabled CPU_FREQ, ACPI_PROCESSOR and INTEL_IDLE as suggested.
>Later I tried also to disable CPU_IDLE, 7300_IDLE (intel chipset idle memory) and all power efficiency settings in the BIOS, but without success.
>The experienced behaviour is following:
>when the process is waiting for an event, the processor lowers its frequency and when the event happens it tries to restore (slowly) its nominal frequency.
>
>Do you have any advice or tip what to try, because I'm really desperate?
>
>Best regards
>
>Petr Cervenka
>
>configuration:
>Xeon E3-1220 (Sandy Bridge)
>linux-2.6.38.8
>xenomai-2.5.6
>

Re: [Xenomai-help] Frequency downscaling of new intel CPU

[Eric Noulard]

2011/10/3 Petr Cervenka <grugh@domain.hid>:
>
> Hello again.
> I know, that my problem is not so close related with the xenomai. But I think many of xenomai users with new intel CPUs have similar problem.
> I tried it it with two different main boards and none of the BIOS settings helped. (C-states: disabled, Speedstep: disabled, CPU idle: high performance, TurboBoost: disabled, ...). My theory is that the linux kernel overrides BIOS settings, but I don't know how to prove it.
> Any ideas that could help me?

Did you verify that you don't have any cpufred related module loaded?

What does

$ lsmod | grep cpu

indicate?

You can try that one too:

$ find /sys/devices/system -name "cpufreq"


-- 
Erk
Membre de l'April - « promouvoir et défendre le logiciel libre » -
http://www.april.org

Re: [Xenomai-help] Frequency downscaling of new intel CPU

[Julien Delange]

On Mon, Oct 3, 2011 at 4:39 AM, Petr Cervenka <grugh@domain.hid> wrote:
> Any ideas that could help me?

What about looking at /proc/cpuinfo while running your system ? If you
see that the speed is not the maximal, indeed, the frequency scaling
system changes the speed of the processor.

Re: [Xenomai-help] Frequency downscaling of new intel CPU

[Petr Červenka]

I'm sure that CPU_FREQ, ACPI_PROCESSOR, CPU_IDLE, INTEL_IDLE and 
7300_IDLE were disabled in kernel configuration and no cpufreq module is 
loaded (or even created).

Later I tried (desperately) to enable most of these settings and 
although I set performance governor to cpufreq, no measurable change 
happened (using acpi_cpufreq module).
But cpuidle (when compiled in kernel) shows that a lot of time is spent 
in power saving states 
(/sys/devices/system/cpu/cpu<n>/cpuidle/state<i>/..., where <i> is also 
2 and 3).

And on one of the mainboards I tried, the acpi_cpufreq was unable to 
load, perhaps because of the disabled BIOS settings).

I'm now working with windows, so the sysfs path is what I remember 
(maybe not so well).

Petr

Dne 3.10.2011 10:47, Eric Noulard napsal(a):
> 2011/10/3 Petr Cervenka<grugh@domain.hid>:
>> Hello again.
>> I know, that my problem is not so close related with the xenomai. But I think many of xenomai users with new intel CPUs have similar problem.
>> I tried it it with two different main boards and none of the BIOS settings helped. (C-states: disabled, Speedstep: disabled, CPU idle: high performance, TurboBoost: disabled, ...). My theory is that the linux kernel overrides BIOS settings, but I don't know how to prove it.
>> Any ideas that could help me?
> Did you verify that you don't have any cpufred related module loaded?
>
> What does
>
> $ lsmod | grep cpu
>
> indicate?
>
> You can try that one too:
>
> $ find /sys/devices/system -name "cpufreq"
>
>

Re: [Xenomai-help] Frequency downscaling of new intel CPU

[Petr Cervenka]

> On Mon, Oct 3, 2011 at 4:39 AM, Petr Cervenka <grugh@domain.hid> wrote:

> >  Any ideas that could help me?

> What about looking at /proc/cpuinfo while running your system ? If you
> see that the speed is not the maximal, indeed, the frequency scaling
> system changes the speed of the processor.

/proc/cpuinfo shows always full frequency clock (3292.500 MHz)

More details:
1) two xenomai tasks are running with 100us period (controlled by external hardware)
2) one task communicates with hardware (receives data via rtnet) and passes them by rt_queue to the second task (on  another core)
3) second task does some computations and returns processed data to the first task
4) first task sends data to hardware

The problem is that during waiting on new data from hw CPU puts itself in some deeper idle state with lower frequency. A then, when the data are available, it slowly wakes up to normal state.
In other words: computation takes longer time than it takes on older computers (depends on overall CPU load).
When some artificial 100% load is running on another core, the computation is as fast as it should be.

Petr

Re: [Xenomai-help] Frequency downscaling of new intel CPU

[Eric Noulard]

2011/10/4 Petr Cervenka <grugh@domain.hid>:
>
>> On Mon, Oct 3, 2011 at 4:39 AM, Petr Cervenka <grugh@domain.hid> wrote:
>
>> >  Any ideas that could help me?
>
>> What about looking at /proc/cpuinfo while running your system ? If you
>> see that the speed is not the maximal, indeed, the frequency scaling
>> system changes the speed of the processor.
>
> /proc/cpuinfo shows always full frequency clock (3292.500 MHz)
>
> More details:
> 1) two xenomai tasks are running with 100us period (controlled by external hardware)
> 2) one task communicates with hardware (receives data via rtnet) and passes them by rt_queue to the second task (on  another core)
> 3) second task does some computations and returns processed data to the first task
> 4) first task sends data to hardware
>
> The problem is that during waiting on new data from hw CPU puts itself in some deeper idle state with lower frequency. A then, when the data are available, it slowly wakes up to normal state.
> In other words: computation takes longer time than it takes on older computers (depends on overall CPU load).
> When some artificial 100% load is running on another core, the computation is as fast as it should be.

Then may be you should ask your computer provider (for the BIOS main board)
and/or intel representative (for more info. about Sandy Bridge dynamic
frequency features)
whether if it's an expected behavior of this kind of processor to
behave like that?

I think it would be nice to know because you assume the processor is
"slowing down" but this
cannot be stated from usual way monitoring processor frequency.

That said, from this small pieces of informations:
http://www.anandtech.com/show/3922/intels-sandy-bridge-architecture-exposed/7

Sandy Bridge seems to be a deterministic computing nightmare,
unless all those nice features (TurboBoost and the like) could be disabled ...
I mean for sure not theoretically.

May be the BIOS is flawed and it is not disabling the dynamic
frequency features properly?
-- 
Erk
Membre de l'April - « promouvoir et défendre le logiciel libre » -
http://www.april.org

Re: [Xenomai-help] Frequency downscaling of new intel CPU

[Gilles Chanteperdrix]

On 09/26/2011 09:02 AM, Petr Cervenka wrote:
> 
> Hello.
> 
> I recently tried newer computer and I suprusingly realized that it is bigger load than the old one with our xenomai application.
> The problem is in the frequency downscaling of the processor when the load in under 100%.
> Of course I have disabled CPU_FREQ, ACPI_PROCESSOR and INTEL_IDLE as suggested.
> Later I tried also to disable CPU_IDLE, 7300_IDLE (intel chipset idle memory) and all power efficiency settings in the BIOS, but without success.
> The experienced behaviour is following:
> when the process is waiting for an event, the processor lowers its frequency and when the event happens it tries to restore (slowly) its nominal frequency.
> 
> Do you have any advice or tip what to try, because I'm really desperate?

Have you tried:
- enabling SMI workaround (and do not forget to check the boot logs to
check that it actually works with the chipset you use);
- if that fails, adding "idle=poll" to kernel arguments?

-- 
                                                                Gilles.

Re: [Xenomai-help] Frequency downscaling of new intel CPU

[Petr Cervenka]

>Have you tried:
>- enabling SMI workaround (and do not forget to check the boot logs to
>check that it actually works with the chipset you use);
>- if that fails, adding "idle=poll" to kernel arguments?
>

Enabling SMI workaround didn't make any change. What should I look after in boot logs?

Kernel parameter idle=poll displayed thermal warnings and freq. throttling after cca 30 sec. of application run.
But when fan speed was selected to Full speed at BIOS, the machine worked fast as it should be. Quite impressive ;-)
Other idle options were same as none, only idle=mwait frezzes during early boot stage.

Thank you for your help.

Petr

PS: Do you have any other idea without overheating side-efects

Re: [Xenomai-help] Frequency downscaling of new intel CPU

[Gilles Chanteperdrix]

On 10/05/2011 03:01 PM, Petr Cervenka wrote:
> 
>> Have you tried:
>> - enabling SMI workaround (and do not forget to check the boot logs to
>> check that it actually works with the chipset you use);
>> - if that fails, adding "idle=poll" to kernel arguments?
>>
> 
> Enabling SMI workaround didn't make any change. What should I look after in boot logs?

SMI

-- 
					    Gilles.