About dynamic switch to offload/nocb on non-server Intel machines

About dynamic switch to offload/nocb on non-server Intel machines

[Joel Fernandes]

Dear Paul and Frederic,

Len Brown on CC provided a good idea in a meeting on using intel_pstate to
detect if device is a server or a consumer device.

Can we perhaps use this to trigger the offload/nocb mechanisms on consumer
devices / distros? I was thinking the dynamic offload and the "one kernel"
work that Frederic did can be trigger from such detection.

It is true userspace can do the trigger, but maybe it is better to not leave
it to userspace so we'd save power on bad userspaces.

Come to think of it, this can even be done at boot time if we can read the
below state early enough:

> intel_pstate.c demonstrates how to detect a server when ACPI is present:
>
>
>
> #ifdef CONFIG_ACPI
>
>
>
> static bool intel_pstate_acpi_pm_profile_server(void)
>
> {
>
>         if (acpi_gbl_FADT.preferred_profile =3D=3D PM_ENTERPRISE_SERVER |=
|
>
>             acpi_gbl_FADT.preferred_profile =3D=3D PM_PERFORMANCE_SERVER)
>
>                 return true;
>
>
>
>         return false;
>
> }
>

Re: About dynamic switch to offload/nocb on non-server Intel machines

[Alison Chaiken]

Resending to mailng list due to HTML formatting . . . sorry.

-- Alison

---

On Fri, Jul 8, 2022 at 10:19 AM Joel Fernandes <joel@joelfernandes.org> wrote:
>
> Dear Paul and Frederic,
>
> Len Brown on CC provided a good idea in a meeting on using intel_pstate to
> detect if device is a server or a consumer device.


Robots are neither a server nor a consumer device.   Please let's not
force industrial users of RT to choose between two inapt sets of
configuration options.

>
> Can we perhaps use this to trigger the offload/nocb mechanisms on consumer
> devices / distros? I was thinking the dynamic offload and the "one kernel"
> work that Frederic did can be trigger from such detection.
>
> It is true userspace can do the trigger, but maybe it is better to not leave
> it to userspace so we'd save power on bad userspaces.
>
> Come to think of it, this can even be done at boot time if we can read the
> below state early enough:
>
> > intel_pstate.c demonstrates how to detect a server when ACPI is present:
> >
> >
> >
> > #ifdef CONFIG_ACPI
> >
> >
> >
> > static bool intel_pstate_acpi_pm_profile_server(void)
> >
> > {
> >
> >         if (acpi_gbl_FADT.preferred_profile =3D=3D PM_ENTERPRISE_SERVER |=
> |
> >
> >             acpi_gbl_FADT.preferred_profile =3D=3D PM_PERFORMANCE_SERVER)
> >
> >                 return true;
> >
> >
> >
> >         return false;
> >
> > }
> >


Robotics applications often rely on fairly "big iron" that is
configured with CONFIG_CPU_FREQ_GOV_PERFORMANCE=y.   Servers and
robotics are quite different use cases, so setting configuration that
would be useful for servers whenever the CPU_FREQ_GOV_PERFORMANCE is
selected would create a headache for many users.

I know, how about a new CONFIG_ANDROID?*

Best wishes,
Alison Chaiken
Aurora Innovation


*That was a joke.

Re: About dynamic switch to offload/nocb on non-server Intel machines

[Joel Fernandes]

Alison, Nice to hear from you!

> On Fri, Jul 8, 2022 at 10:19 AM Joel Fernandes <joel@joelfernandes.org> wrote:
> >
> > Dear Paul and Frederic,
> >
> > Len Brown on CC provided a good idea in a meeting on using intel_pstate to
> > detect if device is a server or a consumer device.
> 
> 
> Robots are neither a server nor a consumer device.   Please let's not
> force industrial users of RT to choose between two inapt sets of
> configuration options.

I see your point.

> > Can we perhaps use this to trigger the offload/nocb mechanisms on consumer
> > devices / distros? I was thinking the dynamic offload and the "one kernel"
> > work that Frederic did can be trigger from such detection.
> >
> > It is true userspace can do the trigger, but maybe it is better to not leave
> > it to userspace so we'd save power on bad userspaces.
> >
> > Come to think of it, this can even be done at boot time if we can read the
> > below state early enough:
> >
> > > intel_pstate.c demonstrates how to detect a server when ACPI is present:
> > >
> > >
> > >
> > > #ifdef CONFIG_ACPI
> > >
> > >
> > >
> > > static bool intel_pstate_acpi_pm_profile_server(void)
> > >
> > > {
> > >
> > >         if (acpi_gbl_FADT.preferred_profile =3D=3D PM_ENTERPRISE_SERVER |=
> > |
> > >
> > >             acpi_gbl_FADT.preferred_profile =3D=3D PM_PERFORMANCE_SERVER)
> > >
> > >                 return true;
> > >
> > >
> > >
> > >         return false;
> > >
> > > }
> > >
> 
> 
> Robotics applications often rely on fairly "big iron" that is
> configured with CONFIG_CPU_FREQ_GOV_PERFORMANCE=y.   Servers and
> robotics are quite different use cases, so setting configuration that
> would be useful for servers whenever the CPU_FREQ_GOV_PERFORMANCE is
> selected would create a headache for many users.

Fair enough, I had to ask...

thanks,

 - Joel

> Best wishes,
> Alison Chaiken
> Aurora Innovation
> 
> 
> *That was a joke.

Re: About dynamic switch to offload/nocb on non-server Intel machines

[Paul E. McKenney]

On Fri, Jul 08, 2022 at 10:40:52AM -0700, Alison Chaiken wrote:
> Resending to mailng list due to HTML formatting . . . sorry.
> 
> -- Alison
> 
> ---
> 
> On Fri, Jul 8, 2022 at 10:19 AM Joel Fernandes <joel@joelfernandes.org> wrote:
> >
> > Dear Paul and Frederic,
> >
> > Len Brown on CC provided a good idea in a meeting on using intel_pstate to
> > detect if device is a server or a consumer device.
> 
> Robots are neither a server nor a consumer device.   Please let's not
> force industrial users of RT to choose between two inapt sets of
> configuration options.
> 
> > Can we perhaps use this to trigger the offload/nocb mechanisms on consumer
> > devices / distros? I was thinking the dynamic offload and the "one kernel"
> > work that Frederic did can be trigger from such detection.
> >
> > It is true userspace can do the trigger, but maybe it is better to not leave
> > it to userspace so we'd save power on bad userspaces.
> >
> > Come to think of it, this can even be done at boot time if we can read the
> > below state early enough:
> >
> > > intel_pstate.c demonstrates how to detect a server when ACPI is present:
> > >
> > >
> > >
> > > #ifdef CONFIG_ACPI
> > >
> > >
> > >
> > > static bool intel_pstate_acpi_pm_profile_server(void)
> > >
> > > {
> > >
> > >         if (acpi_gbl_FADT.preferred_profile =3D=3D PM_ENTERPRISE_SERVER |=
> > |
> > >
> > >             acpi_gbl_FADT.preferred_profile =3D=3D PM_PERFORMANCE_SERVER)
> > >
> > >                 return true;
> > >
> > >
> > >
> > >         return false;
> > >
> > > }
> > >
> 
> 
> Robotics applications often rely on fairly "big iron" that is
> configured with CONFIG_CPU_FREQ_GOV_PERFORMANCE=y.   Servers and
> robotics are quite different use cases, so setting configuration that
> would be useful for servers whenever the CPU_FREQ_GOV_PERFORMANCE is
> selected would create a headache for many users.
> 
> I know, how about a new CONFIG_ANDROID?*
> 
> Best wishes,
> Alison Chaiken
> Aurora Innovation
> 
> 
> *That was a joke.

I want a CONFIG_DO_THE_RIGHT_THING!

No, I am not volunteering to implement it.  Why do you ask?  ;-)

							Thanx, Paul