CPU isolation question again

CPU isolation question again

[Luming Yu]

Hi there,

I noticed some discussion about CPU isolation which points me to the
patch set (https://lkml.org/lkml/2011/8/15/245). I'm currently
preparing a RFC-patch-set to automatically pick up a few suitable CPUs
to isolate then kick them out of service for a while. We need to
balance between  thermal & power management And overall system
performance during this operation as much as possible. So
software-cpu-online-offline interface could not be a good option to
me. But to make sure I'm not blindly running on a dead-end path, I'd
check with experts here to ensure it makes some sense to isolate CPUs
to this level, and the idea also makes some sense, and the most
important is it's not implemented yet.

Thanks!!

Re: CPU isolation question again

[Frederic Weisbecker]

On Wed, Jun 27, 2012 at 09:22:09PM +0800, Luming Yu wrote:
> Hi there,
> 
> I noticed some discussion about CPU isolation which points me to the
> patch set (https://lkml.org/lkml/2011/8/15/245). I'm currently
> preparing a RFC-patch-set to automatically pick up a few suitable CPUs
> to isolate then kick them out of service for a while. We need to
> balance between  thermal & power management And overall system
> performance during this operation as much as possible. So
> software-cpu-online-offline interface could not be a good option to
> me. But to make sure I'm not blindly running on a dead-end path, I'd
> check with experts here to ensure it makes some sense to isolate CPUs
> to this level, and the idea also makes some sense, and the most
> important is it's not implemented yet.

I don't understand what you are trying to do and how exactly. How do you
plan to do this isolation and how do you want to balance between thermal
and power?

Re: CPU isolation question again

[Luming Yu]

On Tue, Jul 3, 2012 at 7:28 PM, Frederic Weisbecker <fweisbec@gmail.com> wrote:
> On Wed, Jun 27, 2012 at 09:22:09PM +0800, Luming Yu wrote:
>> Hi there,
>>
>> I noticed some discussion about CPU isolation which points me to the
>> patch set (https://lkml.org/lkml/2011/8/15/245). I'm currently
>> preparing a RFC-patch-set to automatically pick up a few suitable CPUs
>> to isolate then kick them out of service for a while. We need to
>> balance between  thermal & power management And overall system
>> performance during this operation as much as possible. So
>> software-cpu-online-offline interface could not be a good option to
>> me. But to make sure I'm not blindly running on a dead-end path, I'd
>> check with experts here to ensure it makes some sense to isolate CPUs
>> to this level, and the idea also makes some sense, and the most
>> important is it's not implemented yet.
>
> I don't understand what you are trying to do and how exactly. How do you
> plan to do this isolation and how do you want to balance between thermal
> and power?

My question could be wrong as the question arose several weeks ago
when I came across
 drivers/acpi/acpi_paid.c which looks like a real user who need to
request system automatically
pick up a few CPU to get them isolated and deactivated. Later on, I
noticed tglx's cpu hot plug re-work.
I realized we could reuse the interface to do isolation and deactivation work.

Of cause, to pick up which ones to isolate and deactivate is another problem.

cc'ed the author and ACPI maintainer of the driver as well as tglx.

thanks!

Re: CPU isolation question again

[Frederic Weisbecker]

On Wed, Jul 04, 2012 at 08:42:29PM +0800, Luming Yu wrote:
> On Tue, Jul 3, 2012 at 7:28 PM, Frederic Weisbecker <fweisbec@gmail.com> wrote:
> > On Wed, Jun 27, 2012 at 09:22:09PM +0800, Luming Yu wrote:
> >> Hi there,
> >>
> >> I noticed some discussion about CPU isolation which points me to the
> >> patch set (https://lkml.org/lkml/2011/8/15/245). I'm currently
> >> preparing a RFC-patch-set to automatically pick up a few suitable CPUs
> >> to isolate then kick them out of service for a while. We need to
> >> balance between  thermal & power management And overall system
> >> performance during this operation as much as possible. So
> >> software-cpu-online-offline interface could not be a good option to
> >> me. But to make sure I'm not blindly running on a dead-end path, I'd
> >> check with experts here to ensure it makes some sense to isolate CPUs
> >> to this level, and the idea also makes some sense, and the most
> >> important is it's not implemented yet.
> >
> > I don't understand what you are trying to do and how exactly. How do you
> > plan to do this isolation and how do you want to balance between thermal
> > and power?
> 
> My question could be wrong as the question arose several weeks ago
> when I came across
>  drivers/acpi/acpi_paid.c which looks like a real user who need to
> request system automatically
> pick up a few CPU to get them isolated and deactivated. Later on, I
> noticed tglx's cpu hot plug re-work.
> I realized we could reuse the interface to do isolation and deactivation work.
> 
> Of cause, to pick up which ones to isolate and deactivate is another problem.
> 
> cc'ed the author and ACPI maintainer of the driver as well as tglx.

May be I'm confused because we both have our own definition of isolation.
I'm not sure what kind of CPU isolation you're looking for.

Re: CPU isolation question again

[Luming Yu]

On Wed, Jul 4, 2012 at 9:25 PM, Frederic Weisbecker <fweisbec@gmail.com> wrote:
> On Wed, Jul 04, 2012 at 08:42:29PM +0800, Luming Yu wrote:
>> On Tue, Jul 3, 2012 at 7:28 PM, Frederic Weisbecker <fweisbec@gmail.com> wrote:
>> > On Wed, Jun 27, 2012 at 09:22:09PM +0800, Luming Yu wrote:
>> >> Hi there,
>> >>
>> >> I noticed some discussion about CPU isolation which points me to the
>> >> patch set (https://lkml.org/lkml/2011/8/15/245). I'm currently
>> >> preparing a RFC-patch-set to automatically pick up a few suitable CPUs
>> >> to isolate then kick them out of service for a while. We need to
>> >> balance between  thermal & power management And overall system
>> >> performance during this operation as much as possible. So
>> >> software-cpu-online-offline interface could not be a good option to
>> >> me. But to make sure I'm not blindly running on a dead-end path, I'd
>> >> check with experts here to ensure it makes some sense to isolate CPUs
>> >> to this level, and the idea also makes some sense, and the most
>> >> important is it's not implemented yet.
>> >
>> > I don't understand what you are trying to do and how exactly. How do you
>> > plan to do this isolation and how do you want to balance between thermal
>> > and power?
>>
>> My question could be wrong as the question arose several weeks ago
>> when I came across
>>  drivers/acpi/acpi_paid.c which looks like a real user who need to
>> request system automatically
>> pick up a few CPU to get them isolated and deactivated. Later on, I
>> noticed tglx's cpu hot plug re-work.
>> I realized we could reuse the interface to do isolation and deactivation work.
>>
>> Of cause, to pick up which ones to isolate and deactivate is another problem.
>>
>> cc'ed the author and ACPI maintainer of the driver as well as tglx.
>
> May be I'm confused because we both have our own definition of isolation.
> I'm not sure what kind of CPU isolation you're looking for.

At first, it needs not avaiable to scheduler.  Then, it needs in
deepest power saving mode.
At last, it needs available to scheduler again on demand.
Sounds very like a typical soft offline cpu, but needs to be low light weight.

Re: CPU isolation question again

[Frederic Weisbecker]

On Wed, Jul 04, 2012 at 10:12:43PM +0800, Luming Yu wrote:
> On Wed, Jul 4, 2012 at 9:25 PM, Frederic Weisbecker <fweisbec@gmail.com> wrote:
> > On Wed, Jul 04, 2012 at 08:42:29PM +0800, Luming Yu wrote:
> >> On Tue, Jul 3, 2012 at 7:28 PM, Frederic Weisbecker <fweisbec@gmail.com> wrote:
> >> > On Wed, Jun 27, 2012 at 09:22:09PM +0800, Luming Yu wrote:
> >> >> Hi there,
> >> >>
> >> >> I noticed some discussion about CPU isolation which points me to the
> >> >> patch set (https://lkml.org/lkml/2011/8/15/245). I'm currently
> >> >> preparing a RFC-patch-set to automatically pick up a few suitable CPUs
> >> >> to isolate then kick them out of service for a while. We need to
> >> >> balance between  thermal & power management And overall system
> >> >> performance during this operation as much as possible. So
> >> >> software-cpu-online-offline interface could not be a good option to
> >> >> me. But to make sure I'm not blindly running on a dead-end path, I'd
> >> >> check with experts here to ensure it makes some sense to isolate CPUs
> >> >> to this level, and the idea also makes some sense, and the most
> >> >> important is it's not implemented yet.
> >> >
> >> > I don't understand what you are trying to do and how exactly. How do you
> >> > plan to do this isolation and how do you want to balance between thermal
> >> > and power?
> >>
> >> My question could be wrong as the question arose several weeks ago
> >> when I came across
> >>  drivers/acpi/acpi_paid.c which looks like a real user who need to
> >> request system automatically
> >> pick up a few CPU to get them isolated and deactivated. Later on, I
> >> noticed tglx's cpu hot plug re-work.
> >> I realized we could reuse the interface to do isolation and deactivation work.
> >>
> >> Of cause, to pick up which ones to isolate and deactivate is another problem.
> >>
> >> cc'ed the author and ACPI maintainer of the driver as well as tglx.
> >
> > May be I'm confused because we both have our own definition of isolation.
> > I'm not sure what kind of CPU isolation you're looking for.
> 
> At first, it needs not avaiable to scheduler.  Then, it needs in
> deepest power saving mode.
> At last, it needs available to scheduler again on demand.
> Sounds very like a typical soft offline cpu, but needs to be low light weight.

I see. So indeed the latest developments made in CPU hotplug could make it a solution
for you.

Re: CPU isolation question again

[Luming Yu]

On Wed, Jul 4, 2012 at 10:36 AM, Frederic Weisbecker <fweisbec@gmail.com> wrote:
> On Wed, Jul 04, 2012 at 10:12:43PM +0800, Luming Yu wrote:
>> On Wed, Jul 4, 2012 at 9:25 PM, Frederic Weisbecker <fweisbec@gmail.com> wrote:
>> > On Wed, Jul 04, 2012 at 08:42:29PM +0800, Luming Yu wrote:
>> >> On Tue, Jul 3, 2012 at 7:28 PM, Frederic Weisbecker <fweisbec@gmail.com> wrote:
>> >> > On Wed, Jun 27, 2012 at 09:22:09PM +0800, Luming Yu wrote:
>> >> >> Hi there,
>> >> >>
>> >> >> I noticed some discussion about CPU isolation which points me to the
>> >> >> patch set (https://lkml.org/lkml/2011/8/15/245). I'm currently
>> >> >> preparing a RFC-patch-set to automatically pick up a few suitable CPUs
>> >> >> to isolate then kick them out of service for a while. We need to
>> >> >> balance between  thermal & power management And overall system
>> >> >> performance during this operation as much as possible. So
>> >> >> software-cpu-online-offline interface could not be a good option to
>> >> >> me. But to make sure I'm not blindly running on a dead-end path, I'd
>> >> >> check with experts here to ensure it makes some sense to isolate CPUs
>> >> >> to this level, and the idea also makes some sense, and the most
>> >> >> important is it's not implemented yet.
>> >> >
>> >> > I don't understand what you are trying to do and how exactly. How do you
>> >> > plan to do this isolation and how do you want to balance between thermal
>> >> > and power?
>> >>
>> >> My question could be wrong as the question arose several weeks ago
>> >> when I came across
>> >>  drivers/acpi/acpi_paid.c which looks like a real user who need to
>> >> request system automatically
>> >> pick up a few CPU to get them isolated and deactivated. Later on, I
>> >> noticed tglx's cpu hot plug re-work.
>> >> I realized we could reuse the interface to do isolation and deactivation work.
>> >>
>> >> Of cause, to pick up which ones to isolate and deactivate is another problem.
>> >>
>> >> cc'ed the author and ACPI maintainer of the driver as well as tglx.
>> >
>> > May be I'm confused because we both have our own definition of isolation.
>> > I'm not sure what kind of CPU isolation you're looking for.
>>
>> At first, it needs not avaiable to scheduler.  Then, it needs in
>> deepest power saving mode.
>> At last, it needs available to scheduler again on demand.
>> Sounds very like a typical soft offline cpu, but needs to be low light weight.
>
> I see. So indeed the latest developments made in CPU hotplug could make it a solution
> for you.

I hope it works as it can solve half of my question if the interface
is light enough.  Another half is about a method to tap which set of
logical processors to isolate. We could leave the question to Admin,
or we could automatically sort it out from an ordered list. Not sure
how many type of cpu set we can find from existing APIs..