Messages in this thread |  | | | Date | Sat, 05 Jan 2008 11:27:21 -0500 | | From | Mark Lord <> | | Subject | Re: + restore-missing-sysfs-max_cstate-attr.patch added to -mm tree |
Pallipadi, Venkatesh wrote:
>> -----Original Message-----
>> From: Mark Lord [mailto:lkml@rtr.ca]
..
>> Okay, with !CONFIG_CPU_IDLE, this works fine -- same as 2.6.23
>> and earlier.
>>
>
> Good to know. Atleast we do not have a regression for 2.6.24 now.
..
Agreed. We're happy here, for now.
>>> Meanwhile, can you give a short summary of how behaviour differs
>>> between CONFIG_CPU_IDLE and !CONFIG_CPU_IDLE ??
>>>
>>> I'm not at all clear on how this really affects things.
>
> With CPU_IDLE, the C-state policy is removed from acpi driver. Ideally
> policy should have nothing to do with ACPI, as ACPI only provides the
> C-state mechanisms. So, with CPU_IDLE, it is not easy to control this
> variable through a acpi driver module at run time. Also, the latency
> interface that was mentioned before is to serve the same purpose in a
> more clear manner (based on the wakeup latency) instead of a C-state
> number which may not mean much from the end user point of view.
>
> I will look at why latency does not work on a single core system
> soon(Was that with UP kernel or SMP kernel?). That way we will have a
> proper cover for this with CPU_IDLE in future.
..
That was with a UP kernel on a UP box.
The latency thingie really seemed to have little or no effect,
whereas setting max_cstate=1 has a quite noticeable positive impact.
Things seemed okay (with the latency thingie) on the SMP machine,
but with two cores it is probably simply more forgiving.
cheers
|  |