cpufreq and changing driver

[Nate Lawson]

Marco Calviani wrote:
> Hi list,
> 
> 2005/12/2, Bruno Ducrot <ducrot at poupinou.org>:
> 
> 
> 
>>>I don't see why you can't run powerd more frequently, I do.. Unless your ACPI
>>>has a problem that means the transition is slow.
>>
>>I'm sure this could not be done under Linux without a lot of
>>problems (it is required to use the /proc things and it's too slow in
>>that case).
>>
>>
>>>I can't imagine that doing 5 (or even 50) syscalls a second is a big CPU load
>>>unless there is a specific problem with sysctls or the cpufreq
>>>infrastructure.
>>
>>If that's possible being not so intrusive with, say 50 syscalls under FreeBSD,
>>then all I said above is indeed stupid crap.
>>
>>
>>>I run powerd like this ->
>>>/usr/sbin/powerd -i 90 -r 30 -a adaptive -b adaptive -n adaptive -p 200
>>>
> 
> 
> Well, i've tried decreasing the polling interval, but there is an
> increased powerd cpu load: at 100ms polling interval the cpu load is
> to an astonishing 20% circa, which i think it's too much for a normal
> use. The sampling rate with ondemand governor in linux kernel is 10ms
> but cpufreqd is at 0.6% on average cpu load.

powerd is not intended to do high speed polling.  If you do that, your 
system will almost never be idle and so we can't save power via Cx.

We don't need high speed sampling right now, we need a predictive 
algorithm.  So until someone implements this, it's moot.

-- 
Nate