Message-ID: <51BE8213.4000701@linux.vnet.ibm.com>
Date: Mon, 17 Jun 2013 11:27:15 +0800
From: Michael Wang <wangyun@linux.vnet.ibm.com>
User-Agent: Mozilla/5.0 (X11; Linux i686; rv:16.0) Gecko/20121011 Thunderbird/16.0.1
MIME-Version: 1.0
To: Lei Wen <leiwen@marvell.com>
CC: Peter Zijlstra <peterz@infradead.org>, Ingo Molnar <mingo@elte.hu>,
        mingo@redhat.com, linux-kernel@vger.kernel.org
Subject: Re: [PATCH] sched: add heuristic logic to pick idle peers
References: <1371435692-18831-1-git-send-email-leiwen@marvell.com>
In-Reply-To: <1371435692-18831-1-git-send-email-leiwen@marvell.com>
Content-Type: text/plain; charset=ISO-8859-1
Content-Transfer-Encoding: 7bit
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 3433
Lines: 98

Hi, Lei

On 06/17/2013 10:21 AM, Lei Wen wrote:
> nr_busy_cpus in sched_group_power structure cannot present the purpose
> for judging below statement:
> "this cpu's scheduler group has multiple busy cpu's exceeding
>  the group's power."
> 
> But only could tell how many cpus is doing their jobs for currently.

AFAIK, this nr_busy_cpus presents how many cpus in local group are not
idle, the logical here in nohz_kick_needed() is:

	if domain cpus share resources and at least 2 cpus in
	local group are not idle, prefer to do balance.

And the idea behind is, we catch the timing when there are idle-cpu and
busy-group and task-moving may cost low.

Your change will remove this timing for balance, I think you may need
some test to prove that this patch will make things better.

Regards,
Michael Wang

> 
> However, the original purpose to add this logic still looks good.
> So we move this kind of logic to find_new_ilb, so that we could pick
> out peer from our sharing resource domain whenever possible.
> 
> Signed-off-by: Lei Wen <leiwen@marvell.com>
> ---
>  kernel/sched/fair.c |   28 ++++++++++++++++++++++------
>  1 file changed, 22 insertions(+), 6 deletions(-)
> 
> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index c61a614..64f9120 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -5368,10 +5368,31 @@ static struct {
>  	unsigned long next_balance;     /* in jiffy units */
>  } nohz ____cacheline_aligned;
> 
> +/*
> + * Add the heuristic logic to try waking up idle cpu from
> + * those peers who share resources with us, so that the
> + * cost would be brought to minimum.
> + */
>  static inline int find_new_ilb(int call_cpu)
>  {
> -	int ilb = cpumask_first(nohz.idle_cpus_mask);
> +	int ilb = nr_cpu_ids;
> +	struct sched_domain *sd;
> +
> +	rcu_read_lock();
> +	for_each_domain(call_cpu, sd) {
> +		/* We loop till sched_domain no longer share resource */
> +		if (!(sd->flags & SD_SHARE_PKG_RESOURCES)) {
> +			ilb = cpumask_first(nohz.idle_cpus_mask);
> +			break;
> +		}
> 
> +		/* else, we would try to pick the idle cpu from peers first */
> +		ilb = cpumask_first_and(nohz.idle_cpus_mask,
> +				sched_domain_span(sd));
> +		if (ilb < nr_cpu_ids)
> +			break;
> +	}
> +	rcu_read_unlock();
>  	if (ilb < nr_cpu_ids && idle_cpu(ilb))
>  		return ilb;
> 
> @@ -5620,8 +5641,6 @@ end:
>   * Current heuristic for kicking the idle load balancer in the presence
>   * of an idle cpu is the system.
>   *   - This rq has more than one task.
> - *   - At any scheduler domain level, this cpu's scheduler group has multiple
> - *     busy cpu's exceeding the group's power.
>   *   - For SD_ASYM_PACKING, if the lower numbered cpu's in the scheduler
>   *     domain span are idle.
>   */
> @@ -5659,9 +5678,6 @@ static inline int nohz_kick_needed(struct rq *rq, int cpu)
>  		struct sched_group_power *sgp = sg->sgp;
>  		int nr_busy = atomic_read(&sgp->nr_busy_cpus);
> 
> -		if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
> -			goto need_kick_unlock;
> -
>  		if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
>  		    && (cpumask_first_and(nohz.idle_cpus_mask,
>  					  sched_domain_span(sd)) < cpu))
> 

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/