Hi Eric,
It depends on how we position the momery power saving
functionality: a feature of Intel MCH or a system level power management
mechanism like cpu power management.
If we position it as a feature of Intel MCH, it could be
implemented in almost OS transparent ways. OS only needs to provide some
hook points in idle thread and interrupt handling framework, MCH driver
will handle all other things including idleness prediction based on
system status, enabling/disabling memory power saving mechanism when all
cpus are idle, etc.
If we promote it to system level as a counterpart of cpupm, it
could be much flexible. OS will take the responsibilty to predict memory
idleness and make decision to enter/leave pwoer saving state, MCH only
acts as a backend to enable/disable memory power saving hardware. This
solution may ease supporting of supporting of NUMA and VM(virtual
machine, in previous mail I took it as Virtual Memory, sorry for that).
You pointed out an interesting point, putting unused or rarely
used memory into power saving state. In current Intel UMA architeture,
we can only put memory into low power state as a whole. But things may
change with future NUMA architecture, it sounds very fantastic to
cooperate with MPO to improve power efficiency.
Eric Saxe <mailto:eric.saxe at sun.com> wrote:
> Hi Gerry,
>> I think this power saving mechanism has little impacts on VM
>> because it will be enabled only when system is idle and no other
>> effects other than reduced memory bandwidth will be visible to OS.
>>
> This sounds interesting. It sounds like this is something that the OS
> might want to know about (in terms of observing and controlling it
> being enabled/disabled) since there could be a performance
> impact...is that right? For example, to strike a good balance between
> power efficiency and performance (bandwidth), the OS could enable the
> feature only when it becomes idle...of it could could enable the
> feature only for memory that isn't (or is lightly) utilized. Might
> that be a consideration?
>
> Thanks!
> -Eric
