Re: [PATCH v3 5/6] sched/fair: Consider SMT in ASYM_PACKING load balance

[Joel Fernandes]

On Thu, May 13, 2021 at 08:49:08AM -0700, Ricardo Neri wrote:
> When deciding to pull tasks in ASYM_PACKING, it is necessary not only to
> check for the idle state of the destination CPU, dst_cpu, but also of
> its SMT siblings.
> 
> If dst_cpu is idle but its SMT siblings are busy, performance suffers
> if it pulls tasks from a medium priority CPU that does not have SMT
> siblings.
> 
> Implement asym_can_pull_tasks() to inspect the state of the SMT siblings
> of both dst_cpu and the CPUs in the candidate busiest group.
> 
> To keep the legacy behavior, add an arch_asym_check_smt_siblings() to
> skip these new checks by default. Individual architectures must override
> the mentioned function to use asym_can_pull_tasks().
> 
> Cc: Aubrey Li <aubrey.li@xxxxxxxxx>
> Cc: Ben Segall <bsegall@xxxxxxxxxx>
> Cc: Daniel Bristot de Oliveira <bristot@xxxxxxxxxx>
> Cc: Dietmar Eggemann <dietmar.eggemann@xxxxxxx>
> Cc: Joel Fernandes (Google) <joel@xxxxxxxxxxxxxxxxx>
> Cc: Mel Gorman <mgorman@xxxxxxx>
> Cc: Quentin Perret <qperret@xxxxxxxxxx>
> Cc: Srinivas Pandruvada <srinivas.pandruvada@xxxxxxxxxxxxxxx>
> Cc: Steven Rostedt <rostedt@xxxxxxxxxxx>
> Cc: Tim Chen <tim.c.chen@xxxxxxxxxxxxxxx>
> Reviewed-by: Len Brown <len.brown@xxxxxxxxx>
> Signed-off-by: Ricardo Neri <ricardo.neri-calderon@xxxxxxxxxxxxxxx>
> ---
> Changes since v2:
>   * Reworded the commit message to reflect updates in code.
>   * Corrected misrepresentation of dst_cpu as the CPU doing the load
>     balancing. (PeterZ)
>   * Removed call to arch_asym_check_smt_siblings() as it is now called in
>     sched_asym().
> 
> Changes since v1:
>   * Don't bailout in update_sd_pick_busiest() if dst_cpu cannot pull
>     tasks. Instead, reclassify the candidate busiest group, as it
>     may still be selected. (PeterZ)
>   * Avoid an expensive and unnecessary call to cpumask_weight() when
>     determining if a sched_group is comprised of SMT siblings.
>     (PeterZ).
> ---
>  include/linux/sched/topology.h |   1 +
>  kernel/sched/fair.c            | 101 +++++++++++++++++++++++++++++++++
>  2 files changed, 102 insertions(+)
> 
> diff --git a/include/linux/sched/topology.h b/include/linux/sched/topology.h
> index 8f0f778b7c91..43bdb8b1e1df 100644
> --- a/include/linux/sched/topology.h
> +++ b/include/linux/sched/topology.h
> @@ -57,6 +57,7 @@ static inline int cpu_numa_flags(void)
>  #endif
>  
>  extern int arch_asym_cpu_priority(int cpu);
> +extern bool arch_asym_check_smt_siblings(void);
>  
>  struct sched_domain_attr {
>  	int relax_domain_level;
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index c8b66a5d593e..3d6cc027e6e6 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -106,6 +106,15 @@ int __weak arch_asym_cpu_priority(int cpu)
>  	return -cpu;
>  }
>  
> +/*
> + * For asym packing, first check the state of SMT siblings before deciding to
> + * pull tasks.
> + */
> +bool __weak arch_asym_check_smt_siblings(void)
> +{
> +	return false;
> +}
> +
>  /*
>   * The margin used when comparing utilization with CPU capacity.
>   *
> @@ -8447,6 +8456,89 @@ group_type group_classify(unsigned int imbalance_pct,
>  	return group_has_spare;
>  }
>  
> +/**
> + * asym_can_pull_tasks - Check whether the load balancing CPU can pull tasks
> + * @dst_cpu:	Destination CPU of the load balancing
> + * @sds:	Load-balancing data with statistics of the local group
> + * @sgs:	Load-balancing statistics of the candidate busiest group
> + * @sg:		The candidate busiet group
> + *
> + * Check the state of the SMT siblings of both @sds::local and @sg and decide
> + * if @dst_cpu can pull tasks. If @dst_cpu does not have SMT siblings, it can
> + * pull tasks if two or more of the SMT siblings of @sg are busy. If only one
> + * CPU in @sg is busy, pull tasks only if @dst_cpu has higher priority.
> + *
> + * If both @dst_cpu and @sg have SMT siblings. Even the number of idle CPUs
> + * between @sds::local and @sg. Thus, pull tasks from @sg if the difference
> + * between the number of busy CPUs is 2 or more. If the difference is of 1,
> + * only pull if @dst_cpu has higher priority. If @sg does not have SMT siblings
> + * only pull tasks if all of the SMT siblings of @dst_cpu are idle and @sg
> + * has lower priority.
> + */
> +static bool asym_can_pull_tasks(int dst_cpu, struct sd_lb_stats *sds,
> +				struct sg_lb_stats *sgs, struct sched_group *sg)
> +{
> +#ifdef CONFIG_SCHED_SMT
> +	int cpu, local_busy_cpus, sg_busy_cpus;
> +	bool local_is_smt, sg_is_smt;
> +
> +	cpu = group_first_cpu(sg);
> +	local_is_smt = sds->local->flags & SD_SHARE_CPUCAPACITY;
> +	sg_is_smt = sg->flags & SD_SHARE_CPUCAPACITY;
> +
> +	sg_busy_cpus = sgs->group_weight - sgs->idle_cpus;
> +
> +	if (!local_is_smt) {
> +		/*
> +		 * If we are here, @dst_cpu is idle and does not have SMT
> +		 * siblings. Pull tasks if candidate group has two or more
> +		 * busy CPUs.
> +		 */
> +		if (sg_is_smt && sg_busy_cpus >= 2)
> +			return true;
> +
> +		/*
> +		 * @dst_cpu does not have SMT siblings. @sg may have SMT
> +		 * siblings and only one is busy. In such case, @dst_cpu
> +		 * can help if it has higher priority and is idle.
> +		 */
> +		return !sds->local_stat.group_util &&
> +		       sched_asym_prefer(dst_cpu, sg->asym_prefer_cpu);
> +	}
> +
> +	/* @dst_cpu has SMT siblings. */
> +
> +	local_busy_cpus = sds->local->group_weight - sds->local_stat.idle_cpus;
> +
> +	if (sg_is_smt) {
> +		int busy_cpus_delta = sg_busy_cpus - local_busy_cpus;
> +
> +		/* Local can always help to even the number busy CPUs. */
> +		if (busy_cpus_delta >= 2)
> +			return true;
> +
> +		if (busy_cpus_delta == 1)
> +			return sched_asym_prefer(dst_cpu,
> +						 sg->asym_prefer_cpu);
> +
> +		return false;
> +	}
> +
> +	/*
> +	 * @sg does not have SMT siblings. Ensure that @sds::local does not end
> +	 * up with more than one busy SMT sibling and only pull tasks if there
> +	 * are not busy CPUs. As CPUs move in and out of idle state frequently,
> +	 * also check the group utilization to smoother the decision.

nit: s/smoother/smoothen/

> +	 */
> +	if (!local_busy_cpus && !sds->local_stat.group_util)
> +		return sched_asym_prefer(dst_cpu, sg->asym_prefer_cpu);

Hmm, I am not sure but is it possible that there some local_busy_cpus yet
group_util is 0? If not just check for !group_util ?

FWIW on the series:
Reviewed-by: Joel Fernandes (Google) <joel@xxxxxxxxxxxxxxxxx>

thanks,

 - Joel

> +
> +	return false;
> +#else
> +	return true;
> +#endif
> +}
> +
>  static inline bool
>  sched_asym(struct lb_env *env, struct sd_lb_stats *sds,  struct sg_lb_stats *sgs,
>  	   struct sched_group *group)
> @@ -8458,6 +8550,9 @@ sched_asym(struct lb_env *env, struct sd_lb_stats *sds,  struct sg_lb_stats *sgs
>  	if (group == sds->local)
>  		return false;
>  
> +	if (arch_asym_check_smt_siblings())
> +		return asym_can_pull_tasks(env->dst_cpu, sds, sgs, group);
> +
>  	return sched_asym_prefer(env->dst_cpu, group->asym_prefer_cpu);
>  }
>  
> @@ -9463,6 +9558,12 @@ static struct rq *find_busiest_queue(struct lb_env *env,
>  		    nr_running == 1)
>  			continue;
>  
> +		/* Make sure we only pull tasks from a CPU of lower priority */
> +		if ((env->sd->flags & SD_ASYM_PACKING) &&
> +		    sched_asym_prefer(i, env->dst_cpu) &&
> +		    nr_running == 1)
> +			continue;
> +
>  		switch (env->migration_type) {
>  		case migrate_load:
>  			/*
> -- 
> 2.17.1
>