Re: [PATCH RFC 0/2] sched: Introduce idle SMT priority for asymmetric capacity systems

[Shrikanth Hegde <sshegde@linux.ibm.com>]

Hi Mete.

On 10/13/26 4:20 PM, Mete Durlu wrote:
> On 12/10/2026 16:49, Vincent Guittot wrote:
>> On Mon, 12 Oct 2026 at 13:04, Mete Durlu <meted@linux.ibm.com> wrote:
>>>
>>> On 09/10/2026 17:32, Vincent Guittot wrote:
>>>> On Thu, 8 Oct 2026 at 17:25, Mete Durlu <meted@linux.ibm.com> wrote:
>>>>>
>>>>> On 08/10/2026 16:58, Shrikanth Hegde wrote:
>>>>>> Hi Meter/Andrea.
>>>>>>
>>>>>> On 10/8/26 5:40 PM, Andrea Righi wrote:
>>>>>>> Hi Mete,
>>>>>>>
>>>>>>> On Thu, Oct 08, 2026 at 10:30:46AM +0200, Mete Durlu wrote:
>>>>>>>> Summary
>>>>>>>> ===========================================================================
>>>>>>>> On systems with asymmetric CPU capacities, the scheduler prefers
>>>>>>>> fully idle
>>>>>>>> cores over idle SMT siblings of busy cores. This works generally well
>>>>>>>> but
>>>>>>>> virtualized platforms where low capacity cores should be avoided are not
>>>>>>>> considered. Introduce a new config option and arch hook to prioritize
>>>>>>>> SMT utilization.
>>>>>>
>>>>>> That's true only under physical CPU contention right? or is it always?
>>>>>
>>>>> Right, because of that s390 treats all CPUs as equal and starts to
>>>>> assign lower capacities to CPUs with low entitlement once a certain
>>>>> steal time threshold is crossed (aka physical CPU contention).
>>>
>>> Hi,
>>> Sorry for causing a misunderstanding. I should have been more patient
>>> and detailed while writing my reply.
>>>
>>>> I'm not sure to get what your topology is because "if s390 treats all
>>>> CPUs as equal"  then you should never use select_idle_capacity()
>>>> because  their capacity should be fixed and equal and
>>>> SD_ASYM_CPUCAPACITY_FULL never set.
>>>
>>> On s390 the entitlement of the vCPUs are signified by an arch specific
>>> value we call polarization. We can tell which vCPUs are assigned high
>>> and low entitlement. This entitlement value tells us which vCPUs are
>>> prone to experience steal time first. Without any contention all
>>> vCPUs behave and perform *almost* the same(of course, we want
>>> some priority for CPUs with high entitlement).
>>>
>>> Prior to experiencing steal time(it's what we use to determine
>>> contention) to get the most out of the system we assign all vCPUs
>>> the same capacities, making the system symmetric like you say.
>>> While there is no contention s390's vCPUs also perform better if
>>> fully idle cores are selected for task placement.
>>>
>>>> If your hypervisor lowers the capacity once a certain threshold is
>>>> crossed, we already account for the steal time in the scheduler with
>>>> irq and remove it from the capacity available for tasks but the
>>>> original capacity doesn't change.
>>> s390's linux implementation has a mechanism called hiperdispatch,
>>> that checks topology changes along with system steal time (and soon
>>> the total free capacity of the pyhsical machine) periodically and
>>> makes vCPU capacity adjustments accordingly.
>>>
>>> After steal time starts ramping up and cross a certain threshold,
>>> we start assigning lower capacities to vCPUs with low entitlement to
>>> hint the scheduler that they should be avoided.
>>
>> If all your vCPUs starts with the same capacity, you should not use
>> asym_capacity or enabled it . What you describe above is system
>> pressure (steal time or other) on your vCPU that change over time but
>> not a capacity change. The value returned by arch_scale_cpu_capacity
>> is not supposed to change at runtime or you need to rebuild the
>> sched_domain topology. As an example, max_allowed_capacity does assume
>> capacity fixed otherwise the misfit detection will sometimes fails.
> 
> Exactly for that reason on s390 we do topology updates after
> capacity updates. Topology changes caused by the underlying hypervisor
> or the cpu capacity changes can get reflected without a problem via
> topology rebuilds. Without topology rebuild we wouldn't even get
> the correct SD_ flags to trigger asym_capacity behavior like you say.
> 
>> What you describe really fit into the prefered core that Shrikanth
>> Hegde has developped
> 
> I agree, but Shrikanth's solution can not bring some of the
> behavior changes required. cpu_preferred_mask can help
> direct task placement away from low capacity cores but idle
> core first approach still places tasks to fully idle low
> capacity cores. This patches proposal tries to move tasks
> first to the idle SMT siblings.
>
> Consider the following;
> On a system with 2 high capacity and 4 low capacity cpus.
> steal governor of vCPU backoff mechanism calculates 2 low capacity
> cpus should be taken off from the cpu_preferred_mask to decrease
> experienced steal time. cpu_preferred_mask has 2 high capacity
> and 2 low capacity cpus.


In this scenario, do you still have reasonable steal time or no?

> 
> Current behavior tires to place tasks to fully idle cores first
> even if they are low capacity cores.
> 
>        c0(high) c1(high)  c2(low)  c3(low)  ...
> -------------------------------------------
> SMT 0 task     task      task     task
> SMT 1
> 
> 
> With the proposal, the fully idle core preference is circumvented
> and the most fitting SMT thread is selected.
> 
>        c0(high) c1(high)  c2(low)  c3(low)  ...
> -------------------------------------------
> SMT 0 task     task
> SMT 1 task     task
> 
> This is beneficial because at the time of contention(system pressure)
> low capacity cores could perform worse than SMT siblings of busy
> high capacity cores, even if there is not enough steal time for
> steal governor of Shrikanth's vCPU backoff to mark said low
> capacity cores as non-preferred.
> 

I read this as tunable problem of low,high thresholds to adjust
preferred_mask same as high capacity cores. No?

> Andrea came up with an idea to extend Shrikanth's mechanism further
> by adding a soft affinity to a selected set of CPUs. Task placement
> would first fully load these CPUs before spilling load onto rest
> of the CPUs on cpu_preferred_mask. That would help achieve the
> behavior I describe above if the soft affinity covers high
> capacity CPUs.
>
> The discussion is still on going though. If Shrikanth thinks this is
> a worthwhile concept then I'd be glad to work on the implementation
> and provide the patches.

What i don't understand is.
- if there is no system pressure to use those low capacity cores, why
   not use them if the task util fits its capacity. If the task utilization
   doesn't, it won't land on them anyways today right?