[PATCH] ARM64: Enable multi-core scheduler support by default

[PATCH] ARM64: Enable multi-core scheduler support by default

[Dietmar Eggemann]

Make sure that the task scheduler domain hierarchy is set-up correctly
on systems with single or multi-cluster topology.

Signed-off-by: Dietmar Eggemann <dietmar.eggemann@arm.com>
---
 arch/arm64/configs/defconfig | 1 +
 1 file changed, 1 insertion(+)

diff --git a/arch/arm64/configs/defconfig b/arch/arm64/configs/defconfig
index 34d71dd86781..dc70d715bc38 100644
--- a/arch/arm64/configs/defconfig
+++ b/arch/arm64/configs/defconfig
@@ -51,6 +51,7 @@ CONFIG_PCI=y
 CONFIG_PCI_MSI=y
 CONFIG_PCI_XGENE=y
 CONFIG_SMP=y
+CONFIG_SCHED_MC=y
 CONFIG_PREEMPT=y
 CONFIG_KSM=y
 CONFIG_TRANSPARENT_HUGEPAGE=y
-- 
1.9.1

[PATCH] ARM64: Enable multi-core scheduler support by default

[Punit Agrawal]

Hi Catalin, Will,

Dietmar Eggemann <dietmar.eggemann@arm.com> writes:

> Make sure that the task scheduler domain hierarchy is set-up correctly
> on systems with single or multi-cluster topology.
>
> Signed-off-by: Dietmar Eggemann <dietmar.eggemann@arm.com>

FWIW,
        Acked-by: Punit Agrawal <punit.agrawal@arm.com>

I am not sure how we update the defconfig. Can you pick this up for 4.4?
Or will this be picked up after the merge window?

Thanks,
Punit

> ---
>  arch/arm64/configs/defconfig | 1 +
>  1 file changed, 1 insertion(+)
>
> diff --git a/arch/arm64/configs/defconfig b/arch/arm64/configs/defconfig
> index 34d71dd86781..dc70d715bc38 100644
> --- a/arch/arm64/configs/defconfig
> +++ b/arch/arm64/configs/defconfig
> @@ -51,6 +51,7 @@ CONFIG_PCI=y
>  CONFIG_PCI_MSI=y
>  CONFIG_PCI_XGENE=y
>  CONFIG_SMP=y
> +CONFIG_SCHED_MC=y
>  CONFIG_PREEMPT=y
>  CONFIG_KSM=y
>  CONFIG_TRANSPARENT_HUGEPAGE=y

[PATCH] ARM64: Enable multi-core scheduler support by default

[Catalin Marinas]

On Mon, Oct 19, 2015 at 05:55:49PM +0100, Dietmar Eggemann wrote:
> Make sure that the task scheduler domain hierarchy is set-up correctly
> on systems with single or multi-cluster topology.
> 
> Signed-off-by: Dietmar Eggemann <dietmar.eggemann@arm.com>
> ---
>  arch/arm64/configs/defconfig | 1 +
>  1 file changed, 1 insertion(+)
> 
> diff --git a/arch/arm64/configs/defconfig b/arch/arm64/configs/defconfig
> index 34d71dd86781..dc70d715bc38 100644
> --- a/arch/arm64/configs/defconfig
> +++ b/arch/arm64/configs/defconfig
> @@ -51,6 +51,7 @@ CONFIG_PCI=y
>  CONFIG_PCI_MSI=y
>  CONFIG_PCI_XGENE=y
>  CONFIG_SMP=y
> +CONFIG_SCHED_MC=y
>  CONFIG_PREEMPT=y
>  CONFIG_KSM=y
>  CONFIG_TRANSPARENT_HUGEPAGE=y

Question: does enabling this option have any (positive) effect on the
vanilla kernel or it requires additional patches like EAS?

Thanks.

-- 
Catalin

[PATCH] ARM64: Enable multi-core scheduler support by default

[Dietmar Eggemann]

On 10/29/2015 05:19 PM, Catalin Marinas wrote:
> On Mon, Oct 19, 2015 at 05:55:49PM +0100, Dietmar Eggemann wrote:
>> Make sure that the task scheduler domain hierarchy is set-up correctly
>> on systems with single or multi-cluster topology.
>>
>> Signed-off-by: Dietmar Eggemann <dietmar.eggemann@arm.com>
>> ---
>>   arch/arm64/configs/defconfig | 1 +
>>   1 file changed, 1 insertion(+)
>>
>> diff --git a/arch/arm64/configs/defconfig b/arch/arm64/configs/defconfig
>> index 34d71dd86781..dc70d715bc38 100644
>> --- a/arch/arm64/configs/defconfig
>> +++ b/arch/arm64/configs/defconfig
>> @@ -51,6 +51,7 @@ CONFIG_PCI=y
>>   CONFIG_PCI_MSI=y
>>   CONFIG_PCI_XGENE=y
>>   CONFIG_SMP=y
>> +CONFIG_SCHED_MC=y
>>   CONFIG_PREEMPT=y
>>   CONFIG_KSM=y
>>   CONFIG_TRANSPARENT_HUGEPAGE=y
>
> Question: does enabling this option have any (positive) effect on the
> vanilla kernel or it requires additional patches like EAS?

It improves the task scheduler decisions on multi-cluster systems. You 
have to iterate over all cpu's less frequently when gathering the sched 
group statistics e.g. during CFS load-balancing.

You essentially replace a flat (all cpus) with an hierarchical approach 
(lowest level: sched group eq. cpu, next level: sched group eq. cluster) 
and do CFS load balancing frequently between two cpu's of the same 
cluster (lowest level) and less frequently between two cpus from 
different cluster.

Not sure that I would see an effect on the performance score on a JUNO 
board when running hackbench. Could give it a try ... Probably easier to 
spot on a machine with >32 logical cpus.

EAS has to work with both approaches although we do all our tests with 
CONFIG_SCHED_MC=y. EAS operates on sched groups so the actual structure 
of the sched domain hierarchy is hidden. (at least in theory).

For me CONFIG_SCHED_MC=y is something I would expect when I evaluate the 
task scheduler on a multi-cluster system.

-- Dietmar

[...]

[PATCH] ARM64: Enable multi-core scheduler support by default

[Catalin Marinas]

On Fri, Oct 30, 2015 at 12:29:00PM +0000, Dietmar Eggemann wrote:
> On 10/29/2015 05:19 PM, Catalin Marinas wrote:
> >On Mon, Oct 19, 2015 at 05:55:49PM +0100, Dietmar Eggemann wrote:
> >>Make sure that the task scheduler domain hierarchy is set-up correctly
> >>on systems with single or multi-cluster topology.
> >>
> >>Signed-off-by: Dietmar Eggemann <dietmar.eggemann@arm.com>
> >>---
> >>  arch/arm64/configs/defconfig | 1 +
> >>  1 file changed, 1 insertion(+)
> >>
> >>diff --git a/arch/arm64/configs/defconfig b/arch/arm64/configs/defconfig
> >>index 34d71dd86781..dc70d715bc38 100644
> >>--- a/arch/arm64/configs/defconfig
> >>+++ b/arch/arm64/configs/defconfig
> >>@@ -51,6 +51,7 @@ CONFIG_PCI=y
> >>  CONFIG_PCI_MSI=y
> >>  CONFIG_PCI_XGENE=y
> >>  CONFIG_SMP=y
> >>+CONFIG_SCHED_MC=y
> >>  CONFIG_PREEMPT=y
> >>  CONFIG_KSM=y
> >>  CONFIG_TRANSPARENT_HUGEPAGE=y
> >
> >Question: does enabling this option have any (positive) effect on the
> >vanilla kernel or it requires additional patches like EAS?
> 
> It improves the task scheduler decisions on multi-cluster systems. You have
> to iterate over all cpu's less frequently when gathering the sched group
> statistics e.g. during CFS load-balancing.
> 
> You essentially replace a flat (all cpus) with an hierarchical approach
> (lowest level: sched group eq. cpu, next level: sched group eq. cluster) and
> do CFS load balancing frequently between two cpu's of the same cluster
> (lowest level) and less frequently between two cpus from different cluster.
> 
> Not sure that I would see an effect on the performance score on a JUNO board
> when running hackbench. Could give it a try ... Probably easier to spot on a
> machine with >32 logical cpus.

I did some tests on Juno last week with and without this option and I
haven't noticed any difference.

Anyway, thanks for the explanation. Patch applied.

-- 
Catalin