powerd and increase in energy need
rizzo at iet.unipi.it
Wed Mar 21 17:18:42 UTC 2012
On Wed, Mar 21, 2012 at 09:32:47AM -0700, Kevin Oberman wrote:
> If you are trying to reduce power consumption, why are you limiting Cx
> states to C2 (which save little) and not C3 (which will save a LOT of
> power when the CPU is not heavily loaded).
> If it is due to the system hanging, it is almost certainly because you
> have throttling enabled. Throttling, either by the use of TCC (also
> called P4TCC) or the older, externally implemented throttling
> mechanism, is a BAD BAD THING! I have complained for years about it
> being the default. It is intended for thermal control, not power
> management. The power savings will be negligible and, in combination
> with deep sleep modes (Cx > 2) can and do result in the CPU going into
> deep sleep and never waking up.
> You can (and should) disable them in /boot/loader.conf with:
> # Disable CPU throttling
> This should greatly reduce the large number of "frequencies"
> available, but they will be the ones provided by EST.which really do
> reduce power consumption. (I put frequencies in quotation marks
> because throttling does not really change the clock speed. It simply
> skips 'N' of every 8 clock cycles. Still, compared to C3 and higher,
> EST is a minor power savings.
interesting. Can you elaborate on that ?
This is one of my new machines,
hw.model: Intel(R) Core(TM) i7-2600K CPU @ 3.40GHz
dev.cpu.0.%desc: ACPI CPU
dev.cpu.0.%pnpinfo: _HID=none _UID=0
dev.cpu.0.freq_levels: 3401/255000 3300/245000 3200/236000 3100/227000 3000/218000 2900/209000 2800/200000 2700/192000 2600/183000 2500/175000 2400/167000 2300/159000 2200/151000 2100/143000 1837/125125 1600/87000 1400/76125 1200/65250 1000/54375 800/43500 600/32625 400/21750 200/10875
dev.cpu.0.cx_supported: C1/1 C2/80 C3/104
dev.cpu.0.cx_usage: 100.00% 0.00% 0.00% last 210us
Here using C1 or C3 does not seem to make any difference at all in
terms of power (measured with a Kill-a-watt style wattmeter), and
irrespective of the frequency setting the idle power does not change
much (48..56W across the entire range).
This is on FreeBSD 9 with HZ=1000, maybe some marginal savings can
be achieved setting HZ=100 (at some point i will give it a try).
Under heavy load (16 threads running infinite loops across the 4
cores) the power goes up as expected (up to 118W for 4 cores, 76W
using just 1 core), and powerd seems to do a decent job in keeping
the idle power low.
I guess that the credit for power saving goes mostly to the CPU
architects. Powerd only gives second-order savings, and C1 vs. C3
is ineffective, at least for HZ=1000
CPU Power (watts)
freq idle 16 threads
200 48 51
2200 52 83
3200 54 115
3401 56 118
powerd 48 118
More information about the freebsd-stable