CPU throttling & power conservation expectations

I've been on this list for about a week and have only seen 1 email from a human, wonder if anyone else is out there.

Anyways I had a quick question on using cpufreq to reduce power usage by clocking the CPUs down.

I tested it out on a selection of about a half dozen HP DL320G4s running Pentium D 3Ghz, which appear to be the '925' model.

kernels were primarily 2.6.17-1.2142_FC4smp. I used the p4-clockmod and cpufreq_powersave modules. cpufreq-info reported that the cpus were successfully clocked down to the ~300-400mhz range(forgot exactly what speed).

I expected at least some reduction in power usage, but to my suprise my intelligent power strips did not report even 0.1 amps(@110V) worth of power savings.

I did the same on a selection of HP DL360G4p systems with Xeon 3.0-3.6Ghz, with similar results, no measurable drop in power utilization, despite cpufreq saying that the cpus were clocked down.

I was just wondering if this mirrors other people's experiences or if something is broken.

I do notice of course that power usage goes up if I run high loads on the systems. I was just expecting that clocking a cpu from 3+Ghz down to sub 500mhz would show some measurable amount of power savings, and if not then there doesn't seem to be a lot of point in supporting these CPUs.

We intend to migrate to a newer distribution soon and I can try again on that kernel, but these systems are fairly old as-is(especially the DL360G4s).

thanks!

nate

p4-clockmod throttles the clock frequency but doesn't adjust the processor voltage. You should be using the speedstep-centrino or acpi drivers.

Shouldn't power utilization be measured in watts, not amps? My electrical engineering classes were a long time ago.

I've observed a dual-processor Opteron system pulling 150 watts from the wall at full frequency and voltage, and then dropping to 75 watts at a lower pstate. We did this as a demo at Linux World Expo in 2004.

Also see Paul Devriendt's paper on the same:
http://www.linuxsymposium.org/proceedings/LinuxSymposium2004_V1.pdf

-Mark Langsdorf
Operating System Research Center
AMD, Inc.

ok, that would explain some then, those other drivers wouldn't load, must not be compatible.

[root@prod-admin-2 cpufreq]# modprobe acpi-cpufreq
FATAL: Error inserting acpi_cpufreq (/lib/modules/2.6.17-1.2142_FC4smp/kernel/arch/i386/kernel/cpu/cpufreq/acpi-cpufreq.ko):
Device or resource busy

[root@prod-admin-2 i386]# grep -i acpi /boot/config-2.6.17-1.2142_FC4smp # Power management options (ACPI, APM)
# ACPI (Advanced Configuration and Power Interface) Support CONFIG_ACPI=y
[..]

Well, I think you know what I meant :) it's easy enough to convert amps+volts to watts.. My power strips measure draw in amps. My newer HP gear supports retrieving power usage via the management port, though the versions of the cpufreq drivers in the kernel don't run on those systems either.

thanks a bunch for the quick response I'll check that PDF out too. I'll dig into it more when we go to the newer distribution/kernel.

nate

Does your processors support Enhanced Speedstep feature (does /proc/cpuinfo flags contain 'est')?

If yes, also check whether your BIOS is enabling Enhanced speedstep correctly and whether there are any options in BIOS to enable.

Thanks,
Venki

Looks like my DL320s do (Pentium D), but I don't see the flag on the DL360G4ps (Xeon ~3Ghz). And upon checking I was able to load the acpi-cpufreq module on the DL320 without any errors.

The HP bios isn't big on options, I'll take a look but I don't recall anything talking about speedstep in the bios, though they may have their own terminology for it.

HP does have a 'dynamic power saving' option on the DL320 BIOS (I think it is related to the voltage regulator):

http://h18013.www1.hp.com/products/servers/management/ilo/power-regulator.html

"HP Dynamic Power Savings Mode: Automatically varies processor speed and power usage based on processor utilization. Allows reducing overall power consumption with little or no impact to performance. Does not require OS support."

Which I tried to enable, but the BIOS spit out an error upon reboot saying that option wasn't supported in my specific processor configuration despite their online docs seemingly saying it is supported(running latest bios rev from 11/2006).

I'll run another test on the DL320 test systems with the acpi module loaded and see if it shows a noticable power use reduction.

thanks a bunch!

nate

On Tue, 27 Mar 2007 15:54:30 -0500 Langsdorf, Mark wrote:

[...]

Definitely: adopting the International System of Units (SI), we have

* the watt (symbol: W) is the SI unit of power
* the ampere (symbol: A) is the SI unit of electric current
* the volt (symbol: V) is the SI unit of electric potential difference

See http://www.bipm.fr/ for further details: especially
http://www.bipm.fr/en/si/si_brochure/chapter2/2-1/ampere.html
http://www.bipm.fr/en/si/si_brochure/chapter2/2-2/table3.html

Just in case yours was a serious question... ;-)

 http://frx.netsons.org/doc/nanodocs/etch_workstation_install.html
 Need to read a Debian etch installation walk-through?
..................................................... Francesco Poli .
 GnuPG key fpr == C979 F34B 27CE 5CD8 DC12  31B5 78F4 279B DD6D FCF4

_______________________________________________
Cpufreq mailing list
[email protected]
http://lists.linux.org.uk/mailman/listinfo/cpufreq

This is not too surprising; the p4-clockmod driver is mostly useless. When the processor is idle, most of it is clock gated (effectively 0 MHz) anyway, so reducing the frequency makes no difference. Under high load p4-clockmod should reduce power a little.

Wes Felter - [email protected]