All,

I am looking forward to reduce the TDP for a server planned to be built.
As low as possible shall be best, is AMD cool'n quiet operating with latest 
OpenBSD ?

Regards

[ message continues ]

Depending on what you where looking at, you can reduce the voltages (if
your BIOS has this much control) and this will lower power/heat. I've
done this on PC's with bad HSF in hot temperatures. Though, like over
clocking, it's an art that requires testing, trying and patience to find
the lowest/highest while still being stable

[ message continues ]

Hello,

I think of doing this too.
What I would like to understand is if I will be able to use the frequency
change 1000 / 2000 MHz dynamic load based.

Regards

[ message continues ]

Do you mean change the frequency depending on load on the computer...?
This is very easy in a virtual environment, I am not sure on machine. I
have seen windows software that allows you to change certain options
while in the OS, though weather you could do this in OpenBSD and
dynamically you will need to see if someone else knows the answer. GPU's
are very easy to do this with...certainly doing it manually, but CPU
stuff I'm not so sure...

[ message continues ]

Ok.
I was thinking this is integrated in the core of AMD processor.
Anyway I will see depending on the sunked power if it is necessary to reduce
it further.

Yes, usually the AMD proc use auto reduce of the frequency during standstill
of the OS.

[ message continues ]

The CPU has the ability to lower it's speed but it's the OS that tells it
when to slow down. That's what apm -C tries to do.

I'm using this at home to reduce power $$$. I've reduced the CPU voltage,
and the speed of the integrated GPU (since it's running headless anyway),
put all HDDs on idle timers (IBM/Hitachi drives have some nice powersaving
features) and my multi-TB storage is usually consuming below 100W intake.
Also, apm -C is pure pleasure and gives a significant reduction with my setup.

Note: When running with the lowest multiplier, HDD I/O performance may suffer.
In my case the lowest CPU rate is at 1000MHz and with full I/O load accross 1
or 2 HDDs the CPU load is below the treshold of the apm -C, so it doesn't
speed up. If I switch it manually with apm -H the transfer rate doubles. No
RAID here so we're speaking about 30MB/s with apm -C vs. 60MB/s for apm -H.
Forgot to mention but this is for stuff served over samba meaning there is
some network I/O involved also.


Regards, Daniel.

PS
dmesg below:

OpenBSD 4.6-stable (SQUID_DISKD) #13: Sat Nov 28 14:28:10 CET 2009
    root@pegasus.plan9.homeunix.net:/usr/src/sys/arch/i386/compile/SQUID_DISK
D
cpu0: AMD Sempron(tm) Processor LE-1150 ("AuthenticAMD" 686-class, 256KB L2
cache) 2.01 GHz
cpu0:
FPU,V86,DE,PSE,TSC,MSR,PAE,MCE,CX8,APIC,SEP,MTRR,PGE,MCA,CMOV,PAT,PSE36,CFLUS
H,MMX,FXSR,SSE,SSE2,SSE3,CX16
real mem  = 1003974656 (957MB)
avail mem = 961691648 (917MB)
mainbus0 at root
bios0 at mainbus0: AT/286+ BIOS, date 02/01/08, BIOS32 rev. 0 @ 0xfb7c0,
SMBIOS rev. 2.4 @ 0xf0100 (46 entries)
bios0: vendor Award Software International, Inc. version "F1" date 02/01/2008
bios0: Gigabyte Technology Co., Ltd. GA-MA74GM-S2H
acpi0 at bios0: rev 0
acpi0: tables DSDT FACP SSDT HPET MCFG APIC
acpi0: wakeup devices USB0(S3) USB1(S3) USB2(S3) USB3(S3) USB4(S3) USB5(S3)
USB6(S3) SBAZ(S4) P2P_(S5) PCE2(S4) PCE3(S4) PCE4(S4) PCE5(S4) PCE6(S4)
PCE7(S4) PCE8(S4) PS2K(S5) PCI0(S5)
acpitimer0 at acpi0: 3579545 Hz, 32 ...
[ message continues ]

Hi,
Re your answer, from man page APM(8) :

     -C      Set apmd(8) to cool running performance adjustment mode.  In
this
             mode, when CPU idle time falls below 10%, apm raises hw.setperf
             as much as necessary.  Otherwise when CPU idle time is above
30%,
             apm lowers hw.setperf as much as possible to reduce heat, noise,
             and power consumption.

     -H      Set apmd(8) to manual performance adjustment mode and hw.setperf
             to 100.

I don't understant why you have lower performances after "apm -C" while in my
opinion it should just adjust low / fast in function of the system load
requirement ? Are disk IO not consideredas CPU load ?

Regards.

[ message continues ]

The disk I/O + samba doesn't stress the CPU enough so it does not speed up.
Which is good :) I expect my disk I/O to keep CPU usage low, that's why we
have all that DMA, I/O controllers and stuff.

Bad thing is, with the lowest CPU multiplier, something else is slowed down.
Maybe the communication with the southbridge or the CPU gets very ineffective
in processing data from the bus or something completly different ... At the
moment
apm -C is not tunable and in my case with my hardware it is not sensitive
enough
but I don't care that much since I'm more concerned with power consumption
than raw performance.

Will it be a problem for you?
I have no idea, you should try it with your hardware.

Regards, Daniel.

[ message continues ]

I have a friend who is going to need very low power consumption.
Therefore I'm currently looking into the ways to do so with OpenBSD and more 
or less dedicated hardware.

[ message continues ]

With all this talk about power reduction...I'm going to toss out one
small suggestion:

Get a Wattmeter, and measure...  Don't waste your time speculating.

An ammeter and high school physics V*A=>Watts doesn't cut it for AC
(in general -- a lot of machines are power-factor corrected now so V*A
becomes QUITE useful again, but some have a really big power factor
still...just discovered a P4-vintage machine running a power factor
of 0.65, which surprised the heck out of me.  And if you have no idea
what I'm talking about, just get a good Wattmeter that understands
real AC Wattage, and don't worry about it).

You really need to test assumptions.  I'm quite confident a number of
you are wasting a lot of time on things that Just Don't Matter in the
Big Picture.  Twisting the most knobs may not improve things enough to
justify the downtime you will see.  Spending a lot of money on bizarre
hardware may look laughable when you see how little power some vastly
superior systems draw.  Some slightly older systems which make no
attempt to brag about magic power saving technology use very modest
amounts of power.  Others are gluttonous pigs, of course.

Wait until OpenBSD is fully booted (and the system is idle) before
worrying about what the meter is showing.  Also, look at what your
AVERAGE load will be on the system, and be realistic -- most of the
time, it will probably be completely idle.

I've been quite pleased by the power draw on some PIII-vintage Celeron
i810-chipset machines.  I haven't got around to putting a "full" PIII
in one to see what it "costs" for the additional performance.  I have
theories, but the meter tells.  On the other hand, this 1.8GHz P4 with
two 80G drives, completely normal mini-tower, is drawing less than 50w
(though it is the one with the wicked power-factor -- 75VA).  I paid
$10 for this machine (yes, great deal), it would take a long time for
a Soekris or other specialty machine to save any money for me, IF it
could do what I wanted it to.

Nick.

[ message continues ]

...go here: http://en.wikipedia.org/wiki/Power_factor

[ message continues ]

Even though the cheap ones try to measure real power and not apparent
power, they are often very inaccurate especially at low loads. Watch out
for cheap meters, and if you have multiple similar machines it's not a
bad idea to measure them all at once.

If you don't believe me, buy two cheap meters of the same kind and one of
another kind, if they're within 10% of each other (with a non-resistive
load and low power) I suggest you go buy a lottery ticket. They're still
useful for relative comparisons, of course.

With a machine consuming 100 Watts, switching from a 65% efficient
supply (that's an optimistic guess for a cheap power supply a few years
ago) to a 80% efficient supply will save 28 watts. If you live in a hot
climate and use air conditioning, it's probably worth it in a 24/7
machine. For a machine consuming 50W the savings are probably not worth
the investment. If you are buying a new machine, the price difference
between a crappy and a good power supply is so small it's a no-brainer.
As a bonus the high efficiency supplies run cool and quiet.

[ message continues ]

Hello,

I did. It's consuming some 90 Watts at idle.
Actually, it's an Athlon but the latest Sempron has an even reduced TDP.
My next server will be based on it.
Actually even 70 Watts is a little bit high for my next server given the fact
it will be in an autonomous environment (small wind/solar generators).

Regards

[ message continues ]

Nick is right, numbers should speak.

Parameters at the time of measuring:
MB: GA-MA74GM-S2H rev1.x (no idea about TDP but according to [1] should be
low)
	- unneeded MB components turned off, check my previous dmesg
	- integrated GPU core slowed down from 400MHz to 200MHz
	- Cool'n'quiet enabled in BIOS
CPU: AMD Sempron LE-1150 (TDP 45W)
	- undervolted from 1.2V to 1.00V
	- passive cooling
PSU: Enermax 400W Liberty (no data for this model but it's bigger brothers had
efficiency around 78% for 114W AC [2])
RAM: 1 stick of 1GB DDR2 800MHz
HDD: 3x 1TB, 2x 500G (4 Hitachi and 1 WD Green)
other: 1 low rpm 12cm system fan, no keyboard, no display

I measured my setup couple of times while setting it up. Numbers are from
memory, may not be accurate. Besides, I used a cheap powermeter, meaning the
absolute values are probably off but the deltas could be somewhat trusted:
1. On boot with everything on and no udervolting of CPU the AC wattage was
somewhere around 150W.
2. Undervolting the CPU to 1.00V and playing with the BIOS shaved off some
20-30W.
3. Setting the Hitachi drives to low power idle reduced the power draw for
another 20-30W, at the time the system was idling at 70-90W.
4. Putting the drives into sleep mode got the system under 70W. I'm not using
this atm, had some problems with long delays while waking up the drives.

All of that was without apm -C, basically the CPU was running at full speed.
For testing apm -C I was lazy and relied on temperatures.
1. Temperatures at idle with apm -C
hw.sensors.it0.temp0=28.00 degC
hw.sensors.it0.temp1=33.00 degC

2. Temperatures at idle with apm -H (from memory and different season)
hw.sensors.it0.temp0=34.00 degC
hw.sensors.it0.temp1=38.00 degC
Couple of degrees difference should mean at least couple of W difference.

Regards, Daniel.

Useful links:
47 watt 7 TB server (disks spun down) -
http://www.silentpcreview.com/forums/viewtopic.php?t=57476
TDP list for Intel chipsets ...
[ message continues ]

If absolute raw power is not mandatory, you may have a look at 
Atom-based servers -like 
http://www.supermicro.es/?opcion=contenido&plt=notas&id=137 for example-.

This servers consumption should make a difference when working on 
renovable energy sources.

Regards!


[ message continues ]

You can use apm. It will only save a few watts, but it may reduce the
cooling costs by reducing the heat generated by the CPU. If you have
_many_ machines you can easily reduce the temperature of the server room
by a few degrees C. 

  Advanced Power Management control
  https://calomel.org/apm_control.html

--
   Calomel @ https://calomel.org
   Open Source Research and Reference



[ message continues ]

Is there something one can do to make the system fall into sleep in OpenBSD ?

Thank you.

[ message continues ]

On Sun, 7 Feb 2010 18:44:13 +0100

Suspending has worked for some time on OpenBSD. Have a look at: zzz
But you might want to do your homework about the wakeing up part.

[ message continues ]

if you mean this as general as you wrote it, could you cite
your sources?  suspend has never worked on any of my machines.
or, more precisely, they fall asleep, they just never wake up.

-f
-- 
questions, questions!  does it ever end?!

[ message continues ]

As far as I can see, any modern processor supports several states C0, C1, ...

Is there any way to take advantage of this into OpenBSD in order to reduce the 
system consumption by falling the CPU in C1 state or higher ?

Regards

[ message continues ]

On Sun, 7 Feb 2010 20:41:31 +0100

You got my point.
Exeption are older apm based Thinkpads, like Theo likes to use on
flights.
And i haven't looked into the recent acpi resume improvements.

[ message continues ]