Re: [PATCH RFC v4 3/3] Documentation: arm: define DT idle states bindings

[Lorenzo Pieralisi <lorenzo.pieralisi@arm.com>]

On Mon, Mar 17, 2014 at 06:26:38PM +0000, Antti P Miettinen wrote:
> From: Lorenzo Pieralisi <lorenzo.pieralisi@arm.com>
> > When we make a decision on what idle state to enter all we do, and
> > that's OS agnostic, is predicting (+checking the next event) the next IRQ and
> > see if it is worth entering a state or not. We have to compare it against
> > a baseline, which is the processor being in standbywfi and that's what
> > these bindings define.
> > 
> > I do not understand why you want to define min-residency against the
> > previous shallower state.
> > 
> > What this binding says is: standbywfi is the shallower idle state in
> > power consumption terms. Deeper idle states save more power than
> > standbywfi if the residency in that state is at least min-residency.
> > 
> > I do not see where the problem is to be honest, maybe I need an example.
> > 
> > Thanks!
> > Lorenzo
> 
> Sorry, I should have explained myself more clearly. I've been
> pondering about these issues somewhat lately so I'm perhaps suffering
> from a bit of a tunnel vision.
> 
> In short, when we choose an idle state based on expected idle duration
> we are not comparing wfi against all possible idle states in turn and
> making a decision between wfi and state X. Instead we want to choose
> among all states the one that gives minimum energy for the expected
> idle time. I'll try to elaborate..
> 
> Entering and exiting idle states takes time at nonzero power. To make
> up for this lost energy we indeed want the time in the idle state to
> be sufficiently long to make up for the lost energy. Now the important
> question here is "make up compared to what?".
> 
> The energy over the idle time can be also interpreted as average
> power. When the idle time increases the average power for a state
> approaches the in-state power. A deeper idle state would be a state
> with lower in-state power and longer entry/exit time. Therefore the
> average power for a deeper idle state drops slower as function of idle
> time than the average power for a shallower idle state. If we'd plot
> the average power for a number of idle states as function of idle
> duration, we'd get a set of "constant over idle time plus constant"
> style curves. Average power for state 0 will drop fastest close to the
> in-state power of state 0. Average power for state 1 will drop slower
> and approach the in-state power of state 1, average power for state 2
> will drop even slower and approach the in-state power of state 3.
> 
> To define that the min-residency is the breakeven time against state 0
> means that we are looking at the curves and asking "when does the
> average power for state X cross the average power for state 0?". But
> that would be the guideline for making a decision between state 0 and
> the state in question. Even if average power for state 2 is below
> the average power of state 0 it is not necessarily yet below the
> average power of state 1. To break even against state 1 the idle
> duration needs to be longer.
> 
> Yet another way to look at this: for three states we can define three
> times of interest:
> - t1: the time when state1 breaks even against state0
> - t2: the time when state2 breaks even against state0
> - t3: the time when state2 breaks even against state1
> and t3 would typically be larger than t2.

Now it is crystal clear, and you are absolutely right, sorry for
misunderstanding.

Help me define it then please:

- min-residency-us

"u32 value representing time in microseconds required for the CPU to be in
the idle state to guarantee power savings maximization".

Rather vague (on purpose), if anyone comes up with a better definition please
shout.

Thanks !
Lorenzo