[v2] sched/fair: Fix the decision for load balance

[PATCH v2] sched/fair: Fix the decision for load balance

Posted by Keisuke Nishimura 2 years, 2 months ago

should_we_balance is called for the decision to do load-balancing.
When sched ticks invoke this function, only one CPU should return
true. However, in the current code, two CPUs can return true. The
following situation, where b means busy and i means idle, is an
example, because CPU 0 and CPU 2 return true.

        [0, 1] [2, 3]
         b  b   i  b

This fix checks if there exists an idle CPU with busy sibling(s)
after looking for a CPU on an idle core. If some idle CPUs with busy
siblings are found, just the first one should do load-balancing.

Fixes: b1bfeab9b002 ("sched/fair: Consider the idle state of the whole core for load balance")
Signed-off-by: Keisuke Nishimura <keisuke.nishimura@inria.fr>
---
 kernel/sched/fair.c | 10 +++++++---
 1 file changed, 7 insertions(+), 3 deletions(-)

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 2048138ce54b..69d63fae34f4 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -11079,12 +11079,16 @@ static int should_we_balance(struct lb_env *env)
 			continue;
 		}
 
-		/* Are we the first idle CPU? */
+		/*
+		 * Are we the first idle core in a MC or higher domain
+		 * or the first idle CPU in a SMT domain?
+		 */
 		return cpu == env->dst_cpu;
 	}
 
-	if (idle_smt == env->dst_cpu)
-		return true;
+	/* Are we the first idle CPU with busy siblings? */
+	if (idle_smt != -1)
+		return idle_smt == env->dst_cpu;
 
 	/* Are we the first CPU of this group ? */
 	return group_balance_cpu(sg) == env->dst_cpu;
-- 
2.34.1

Re: [PATCH v2] sched/fair: Fix the decision for load balance

Posted by Chen Yu 2 years, 2 months ago

On 2023-10-30 at 18:29:46 +0100, Keisuke Nishimura wrote:
> should_we_balance is called for the decision to do load-balancing.
> When sched ticks invoke this function, only one CPU should return
> true. However, in the current code, two CPUs can return true. The
> following situation, where b means busy and i means idle, is an
> example, because CPU 0 and CPU 2 return true.
> 
>         [0, 1] [2, 3]
>          b  b   i  b
> 
> This fix checks if there exists an idle CPU with busy sibling(s)
> after looking for a CPU on an idle core. If some idle CPUs with busy
> siblings are found, just the first one should do load-balancing.
> 
> Fixes: b1bfeab9b002 ("sched/fair: Consider the idle state of the whole core for load balance")
> Signed-off-by: Keisuke Nishimura <keisuke.nishimura@inria.fr>
> ---
>  kernel/sched/fair.c | 10 +++++++---
>  1 file changed, 7 insertions(+), 3 deletions(-)
> 
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index 2048138ce54b..69d63fae34f4 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -11079,12 +11079,16 @@ static int should_we_balance(struct lb_env *env)
>  			continue;
>  		}
>  
> -		/* Are we the first idle CPU? */
> +		/*
> +		 * Are we the first idle core in a MC or higher domain

It is possible that the Cluster domain is lower than a MC.
cluser domain: CPUs share the same L2
MC domain: CPUs share the same LLC

 grep . domain*/{name,flags}
domain0/name:CLS
domain1/name:MC
domain2/name:NUMA
domain0/flags:SD_BALANCE_NEWIDLE SD_BALANCE_EXEC SD_BALANCE_FORK SD_WAKE_AFFINE SD_SHARE_PKG_RESOURCES SD_PREFER_SIBLING 
domain1/flags:SD_BALANCE_NEWIDLE SD_BALANCE_EXEC SD_BALANCE_FORK SD_WAKE_AFFINE SD_SHARE_PKG_RESOURCES SD_PREFER_SIBLING 
domain2/flags:SD_BALANCE_NEWIDLE SD_BALANCE_EXEC SD_BALANCE_FORK SD_WAKE_AFFINE SD_SERIALIZE SD_OVERLAP SD_NUMA 

So, maybe:
Are we the first idle core in a non-SMT domain or higher,

thanks,
Chenyu

> +		 * or the first idle CPU in a SMT domain?
> +		 */
>  		return cpu == env->dst_cpu;
>  	}
>  
> -	if (idle_smt == env->dst_cpu)
> -		return true;
> +	/* Are we the first idle CPU with busy siblings? */
> +	if (idle_smt != -1)
> +		return idle_smt == env->dst_cpu;
>  
>  	/* Are we the first CPU of this group ? */
>  	return group_balance_cpu(sg) == env->dst_cpu;
> -- 
> 2.34.1
>

Re: [PATCH v2] sched/fair: Fix the decision for load balance

Posted by Shrikanth Hegde 2 years, 2 months ago


On 10/31/23 11:29 AM, Chen Yu wrote:
> On 2023-10-30 at 18:29:46 +0100, Keisuke Nishimura wrote:
>> should_we_balance is called for the decision to do load-balancing.
>> When sched ticks invoke this function, only one CPU should return
>> true. However, in the current code, two CPUs can return true. The
>> following situation, where b means busy and i means idle, is an
>> example, because CPU 0 and CPU 2 return true.
>>
>>         [0, 1] [2, 3]
>>          b  b   i  b
>>
>> This fix checks if there exists an idle CPU with busy sibling(s)
>> after looking for a CPU on an idle core. If some idle CPUs with busy
>> siblings are found, just the first one should do load-balancing.
>>

As Chen indicated, it would be better to carry reviewed by tags.

>> Fixes: b1bfeab9b002 ("sched/fair: Consider the idle state of the whole core for load balance")
>> Signed-off-by: Keisuke Nishimura <keisuke.nishimura@inria.fr>
>> ---
>>  kernel/sched/fair.c | 10 +++++++---
>>  1 file changed, 7 insertions(+), 3 deletions(-)
>>
>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
>> index 2048138ce54b..69d63fae34f4 100644
>> --- a/kernel/sched/fair.c
>> +++ b/kernel/sched/fair.c
>> @@ -11079,12 +11079,16 @@ static int should_we_balance(struct lb_env *env)
>>  			continue;
>>  		}
>>  
>> -		/* Are we the first idle CPU? */
>> +		/*
>> +		 * Are we the first idle core in a MC or higher domain
> 
> It is possible that the Cluster domain is lower than a MC.
> cluser domain: CPUs share the same L2
> MC domain: CPUs share the same LLC
> 
>  grep . domain*/{name,flags}
> domain0/name:CLS
> domain1/name:MC
> domain2/name:NUMA
> domain0/flags:SD_BALANCE_NEWIDLE SD_BALANCE_EXEC SD_BALANCE_FORK SD_WAKE_AFFINE SD_SHARE_PKG_RESOURCES SD_PREFER_SIBLING 
> domain1/flags:SD_BALANCE_NEWIDLE SD_BALANCE_EXEC SD_BALANCE_FORK SD_WAKE_AFFINE SD_SHARE_PKG_RESOURCES SD_PREFER_SIBLING 
> domain2/flags:SD_BALANCE_NEWIDLE SD_BALANCE_EXEC SD_BALANCE_FORK SD_WAKE_AFFINE SD_SERIALIZE SD_OVERLAP SD_NUMA 
> 
> So, maybe:
> Are we the first idle core in a non-SMT domain or higher,

Yes. That makes sense. Forgot about recent cluster addition.

> 
> thanks,
> Chenyu
> 
>> +		 * or the first idle CPU in a SMT domain?
>> +		 */
>>  		return cpu == env->dst_cpu;
>>  	}
>>  
>> -	if (idle_smt == env->dst_cpu)
>> -		return true;
>> +	/* Are we the first idle CPU with busy siblings? */
>> +	if (idle_smt != -1)
>> +		return idle_smt == env->dst_cpu;
>>  
>>  	/* Are we the first CPU of this group ? */
>>  	return group_balance_cpu(sg) == env->dst_cpu;
>> -- 
>> 2.34.1
>>

Re: [PATCH v2] sched/fair: Fix the decision for load balance

Posted by Chen Yu 2 years, 2 months ago

On 2023-10-31 at 13:59:30 +0800, Chen Yu wrote:
> On 2023-10-30 at 18:29:46 +0100, Keisuke Nishimura wrote:
> > should_we_balance is called for the decision to do load-balancing.
> > When sched ticks invoke this function, only one CPU should return
> > true. However, in the current code, two CPUs can return true. The
> > following situation, where b means busy and i means idle, is an
> > example, because CPU 0 and CPU 2 return true.
> > 
> >         [0, 1] [2, 3]
> >          b  b   i  b
> > 
> > This fix checks if there exists an idle CPU with busy sibling(s)
> > after looking for a CPU on an idle core. If some idle CPUs with busy
> > siblings are found, just the first one should do load-balancing.
> > 
> > Fixes: b1bfeab9b002 ("sched/fair: Consider the idle state of the whole core for load balance")
> > Signed-off-by: Keisuke Nishimura <keisuke.nishimura@inria.fr>
> > ---
> >  kernel/sched/fair.c | 10 +++++++---
> >  1 file changed, 7 insertions(+), 3 deletions(-)
> > 
> > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> > index 2048138ce54b..69d63fae34f4 100644
> > --- a/kernel/sched/fair.c
> > +++ b/kernel/sched/fair.c
> > @@ -11079,12 +11079,16 @@ static int should_we_balance(struct lb_env *env)
> >  			continue;
> >  		}
> >  
> > -		/* Are we the first idle CPU? */
> > +		/*
> > +		 * Are we the first idle core in a MC or higher domain
> 
> It is possible that the Cluster domain is lower than a MC.
> cluser domain: CPUs share the same L2
> MC domain: CPUs share the same LLC
> 
>  grep . domain*/{name,flags}
> domain0/name:CLS
> domain1/name:MC
> domain2/name:NUMA
> domain0/flags:SD_BALANCE_NEWIDLE SD_BALANCE_EXEC SD_BALANCE_FORK SD_WAKE_AFFINE SD_SHARE_PKG_RESOURCES SD_PREFER_SIBLING 
> domain1/flags:SD_BALANCE_NEWIDLE SD_BALANCE_EXEC SD_BALANCE_FORK SD_WAKE_AFFINE SD_SHARE_PKG_RESOURCES SD_PREFER_SIBLING 
> domain2/flags:SD_BALANCE_NEWIDLE SD_BALANCE_EXEC SD_BALANCE_FORK SD_WAKE_AFFINE SD_SERIALIZE SD_OVERLAP SD_NUMA 
> 
> So, maybe:
> Are we the first idle core in a non-SMT domain or higher,
>

I suppose you can also carry the Reviewed-by tags in V1(Shrikanth and Vincent's)
as there is no code change.

Reviewed-by: Chen Yu <yu.c.chen@intel.com>

thanks,
Chenyu