LWN: Comments on "The PowerClamp driver"

The PowerClamp driver

daniexia — Wed, 26 Jun 2013 16:57:37 +0000

solved, after enabling package C state in BIOS.

The PowerClamp driver

daniexia — Tue, 25 Jun 2013 21:22:10 +0000

Hi,

I have a problem with seeing the effect of powerclamp. First, I enabled intel powerclamp driver in my menuconfig. Then, I built a new kernel 3.9.7. After rebooting, I can see the module as /sys/class/thermal/cooling_device13. When I grep it. It shows,

cur_state:-1
max_state:50
type:intel_powerclamp

The cur_state =-1, which means invalid state. Why is this invalid state?

Thanks.

The PowerClamp driver

aidenn0 — Mon, 29 Apr 2013 17:04:03 +0000

This sounds like something I could really use; I have no A/C where I live, and my server tends to overheat if I compile code on it in August; I could pretty easily hook into the thermal info from user-space to clamp down the CPU time when it gets too hot.

The PowerClamp driver

dlang — Mon, 11 Mar 2013 20:28:26 +0000

not really, it takes a long time for the processor to go to sleep and wake up again, long enough that it's frequently better for the processor to idle at high speed rather than go into sleep mode.

The PowerClamp driver

ssam — Mon, 11 Mar 2013 13:06:13 +0000

if i am running a process that is limited by memory bandwidth, wont the CPU be spend lots of time idle, and so reducing the power consumption already? Can the CPU sleep while it is waiting for something to be fetched from RAM?

The PowerClamp driver

dlang — Thu, 07 Mar 2013 23:30:29 +0000

and remember that even memory access is an external event that is significantly slower than the CPU, so if what you are doing is a small amount of computation on a large amount of memory, you may be able to do it in the same amount of time with a much slower CPU.

The PowerClamp driver

dgm — Thu, 07 Mar 2013 22:53:05 +0000

> If you deliberately hobble the workload to use less power, it takes more time

Not if your work depends on external events, that is, IO. And it all depends on the characteristics of the system. If you can do it 10% slower at half the power consumption, it may very well be worth the wait.

The PowerClamp driver

pjones — Thu, 07 Mar 2013 21:57:49 +0000

energy = power * time. If you deliberately hobble the workload to use less power, it takes more time, but also operates less efficiently. This means you'll use more energy.

The PowerClamp driver

redden0t8 — Thu, 13 Dec 2012 15:59:10 +0000

I'm kind of confused... how does reduced power consumption cost you battery life?

Is it because a given workload will take longer to complete, and therefore take more total power by the time its done?

If that's the case, then it's actually workload dependent. Specifically, I'm thinking of playing retro games that insist on taking 100% of the CPU. They needlessly waste power to achieve much greater that 60 fps, even with the CPU clocked to the lowest scaling frequency. Would PowerClamp not increase battery life in this scenario?

The PowerClamp driver

arjan — Wed, 12 Dec 2012 16:55:29 +0000

the article is clear to me, but I can see people thinking this is for "saving power" (see the first comments)...

The PowerClamp driver

corbet — Wed, 12 Dec 2012 16:44:57 +0000

I'm confused...the word "battery" does not appear in the article. Instead I talk about things like temperature regulation. Was something not clear?

The PowerClamp driver

arjan — Wed, 12 Dec 2012 16:40:37 +0000

powerclamp is a building block towards a solution that you describe.
Note that if you need a "near instant" limit, a kernel level solution isn't going to work, you need something much much faster responding.
But if you can deal with "we need over <some hundreds of milliseconds> the average to be below X", and you can measure X.. this driver is what a small (userspace) control agent can use to actually impact the current consumption

The PowerClamp driver

arjan — Wed, 12 Dec 2012 16:37:43 +0000

Note that this is NOT about getting a longer battery life.
In fact, pretty much all tricks to put a limit on temperature/power consumption like this en up costing you battery life.

This is is about limiting either temperature or current use (on laptops temperature matters, in data centers current matters, but also temperature) due to external constraints.

We have a simple userspace app for example that can control a laptop temperature to just below the point that the fan would come on.
(it's very much prototype code at this point.. we're working on getting it more usable than on the one machine we ran it on).

The PowerClamp driver

idupree — Fri, 07 Dec 2012 01:59:59 +0000

Yeah, alas. I'd hoped it was a "[mechanism] to put an overall limit on the amount of power consumed" but it isn't. I was looking at certain external batteries[*] for travel which have a maximum output rating of 4.2A. My laptop (at 19V DC input) is capable of using more than 4.2A (80 W). The power brick it came with has max 6.32A; from my measurements with some of CPU/GPU/screen at full force, this higher capacity is sometimes necessary. If my system tried to draw more power at any moment, bad things might happen (I'm not sure how bad. Battery damage? Laptop shutdown?). I wouldn't buy that battery without a way to ensure the laptop's total power usage was below the limit -- which is a hard problem which this article does not appear to go anywhere near solving.

[*] http://zikko-store.com/product_view.php?id=7

The PowerClamp driver

Jonno — Thu, 06 Dec 2012 13:56:14 +0000

> Can someone with a clue tell me what is really going on?
Linux kernel CPU power management currently beats all other, but the CPU isn't the only thing drawing power. GPU, PCIe, and HDD power management are all areas where Linux could do significantly better. And Linux userspace is typically less power management aware than that of MacOS X, resulting in the Mac simply doing less work while on battery...

The PowerClamp driver

hummassa — Thu, 06 Dec 2012 11:49:00 +0000

My MacBook also has a considerably long battery life *as long as I don't do anything funny with it* (like transcoding movies or compiling a large piece of software like gcc...)

The PowerClamp driver

ras — Thu, 06 Dec 2012 11:00:15 +0000

> The kernel's power management subsystem has become increasingly effective over recent years, to the point that our CPU power management is said to be second to none.

OK, now I am confused.

I always put MacBook's phenomenal battery life to excellent software power management, and by implication the Linux ecosystem had a bit of catching up to do. Admittedly this based purely on observation that MacBook's are nearly identical. For example the MacBook pro's battery capacity is approx 210kJ, a Dell is 180kJ. The other components - CPU, graphics card, memory, disk drive are all the same. Yet the Mac will last 10 hours, and the Dell 4.

Can someone with a clue tell me what is really going on?