[CentOS] Questions on cpu frequency scaling AMD vs. Intel

Sun Aug 3 15:16:25 UTC 2008
Ned Slider <ned at unixmail.co.uk>

Kai Schaetzl wrote:
> Ned Slider wrote on Sun, 03 Aug 2008 15:09:39 +0100:
> 
>> http://www.centos.org/modules/newbb/viewtopic.php?topic_id=15484&forum=37
> 
> Thanks for the URL, see below!
> 
>> Bottom line - the power saving between having frequency scaling enabled 
>> or not was surprisingly small (only 2-3W). It would appear that these 
>> processors are already fairly efficient at idle and scaling down the 
>> frequency adds little to the overall savings that may be obtained.
> 
> I disagree about the reason. I think they are actually not so efficient. At 
> least not if I compare to a low-voltage CPU. 105 W is a lot, latest AMD 
> quad core low-voltage are at 50W. Did you check core temperature in the two 
> scaling states? It makes a huge difference for me on the AMD (which is 
> allowed to drop from 2500 to 1000). It drops from an already low value (30 
> and 22 Celsius) by more than 10 degrees. The second core always shows the 
> lowest temperature (puzzle?) and it goes down to 6-8 (!) Celsius in idle 
> state with 1000.) I think this will also result on some more substantial 
> savings in Watt consumption. Even, if not, a substantially lower 
> temperature like this is good for a long life of all parts, anyway.
> 

I see no difference on temps reported by coretemp for cpuspeed 
enabled/disabled. I *do* see a huge drop in temps between load and idle 
regardless of cpuspeed.

> I read that thread and am puzzled by acpi-cpufreq being loaded on your 
> machine. If I modprobe it I get an error "device busy". Which makes sense 
> to me as cpufreq_ondemand (which loaded automatically) should have already 
> taken over. I see that behavior on all machines, no matter if Intel or AMD. 
>>From my research yesterday it also looks like use of acpi-cpufreq is 
> somewhat "older" and should not be necessary at all for newer CPUs. So, it 
> should be cpufreq_ondemand alone that does the scaling on your machine. Can 
> you confirm that?

I'm not sure of the function of acpi-cpufreq. I do know that it doesn't 
scale back *without* cpufreq_ondemand (cpuspeed). acpi-cpufreq was 
autoloaded in response to enabling C1E and EIST features in the BIOS 
(which one is responsible I don't know as I enabled both together).

> I also wonder if your machine actually scales up. You listed the output in 
> low/idle state. As I wrote I get the same, just at another level (they 
> probably think Xeon's will be active all the time, anyway, so they allow 
> them to drop not so much). Did you check that the frequency actually goes 
> up to 2400 under load?

Yes, the frequency does scale up under load. I tested by launching a 
scientific app that loads all 4 cores at 100%. As fast as I could 
manually start the app and check the freq, it reported at 2.4GHz. I 
don't know at what point or under what load it will scale back up, and 
if scaling is done on a core by core basis, but it does scale back up 
under full load.