On Mon, Oct 14, 2013 at 02:32:34PM +0100, Peter Zijlstra wrote:
On Fri, Oct 11, 2013 at 06:19:10PM +0100, Morten Rasmussen wrote:
Hi,
I have revised the previous power scheduler proposal[1] trying to address as many of the comments as possible. The overall idea was discussed at LPC[2,3]. The revised design has removed the power scheduler and replaced it with a high level power driver interface. An interface that allows the scheduler to query the power driver for information and provide hints to guide power management decisions in the power driver.
The power driver is going to be a unified platform power driver that can replace cpufreq and cpuidle drivers. Generic power policies will be optional helper functions called from the power driver. Platforms may choose to implement their own policies as part of their power driver.
This RFC series prototypes a part of the power driver interface (cpu capacity hints) and shows how they can be used from the scheduler. More extensive use of the power driver hints and queries is left for later. The focus for now is the power driver interface. The patch series includes a power driver/cpufreq governor that can use existing cpufreq drivers as backend. It has been tested (not thoroughly) on ARM TC2. The cpufreq governor power driver implementation is rather horrible, but it illustrates how the power driver interface can be used. Native power drivers is on the todo list.
The power driver interface is still missing quite a few calls to handle: Idle, adding extra information to the sched_domain hierarchy to guide scheduling decisions (packing), and possibly scaling of tracked load to compensate for frequency changes and asymmetric systems (big.LITTLE).
This set is based on 3.11. I have done ARM TC2 testing based on linux-linaro 2013.08[4] to get cpufreq support for TC2.
What I'm missing is a general overview of why what and how.
In particular; how does this proposal lead to power savings. Is there a mathematical model that supports this framework? Something where if you give it a task set with global utilisation < 1 (ie. there's idle time), it results in less power used.
It is not there yet. What I'm proposing here is just the interface with some very simple examples of how they can be used. It is not the complete set of hooks, but a starting point. In the first round of discussions it was clear that it is quite important to find an interface that can work for everyone. To find a good power optimization model we first need to know what information the platform (represented by the power driver) can provide.
With guidance from the power driver about what level performance we can expect from a cpu, and possibly also the power cost, we can make better load-balancing decisions. We can add task packing support and let the platform decide the degree of packing that meets the power/performance target.
Also, how does this proposal deal with cpufreq's fundamental broken approach to SMP? Afaict nothing considers the effect of one cpu upon another -- something which isn't true at all.
If you are referring to not doing anything clever with the affected cpus in the power driver then yes. I doesn't do anything clever at the moment. However, the go_faster/slower() interface would allow the power driver to refuse increasing the frequency if the power cost can't be justified for some reason. Using the power driver interface the scheduler will know this and be able to try a different balance. Spread load to other cpus in the same frequency domain that are less loaded, if possible.
In fact, I don't see anything except a random bunch of hooks without an over-all picture of how to get less power used.
I will follow up with a better description of the overall picture. The slides I linked to are not really self-explaining.
Thanks, Morten