[Ksummit-2013-discuss] [ATTEND] Linux VM Infrastructure to support Memory Power Management

[Srivatsa S. Bhat]

On 07/29/2013 10:28 PM, Luck, Tony wrote:
>> Few things to consider
>> * Running a cpu core to move stuff around may be expensive or cheap, depending on the cpu one picked
> 
> Also need to consider the expected usage pattern in the future.  While "prediction is very difficult,
> especially about the future"[1] there probably are many cases where a smart agent (likely not in
> the kernel) can get things right often enough to excuse the power lost in the cases where it gets
> it wrong.
> 
> At the device end of the spectrum (phones, tablets ... assuming they have memory split across
> two or more power zones and actually need this) - user activity like shutting off the screen may
> be enough of a clue that they don't plan to use the device for at least the next minute or two.
> Combine that with some accelerometer data (put the device into pocket/bag) and perhaps time
> of day (user going to bed) and you might get even better predictions of longer term low activity.
> We get it wrong here where some external event disrupts the pattern (incoming phone call).
> 
> At the cloud server end of the spectrum - the traffic shaping front ends know where they are
> sending work - and might indicate to a few racks that they can expect low load for a while. So
> they should shuffle memory around to make some banks idle.
> 
> 

These are very interesting points! We'll surely find it beneficial to
expose some sort of policy control to userspace, and allow it to influence
how the kernel deals with memory allocations to save power, by giving
appropriate hints.

Thanks Tony!

Regards,
Srivatsa S. Bhat