From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1755458AbaFPPuM (ORCPT ); Mon, 16 Jun 2014 11:50:12 -0400 Received: from mga09.intel.com ([134.134.136.24]:36328 "EHLO mga09.intel.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752834AbaFPPuL (ORCPT ); Mon, 16 Jun 2014 11:50:11 -0400 X-ExtLoop1: 1 X-IronPort-AV: E=Sophos;i="5.01,487,1400050800"; d="scan'208";a="558300068" Subject: Re: [PATCH] sched: Fast idling of CPU when system is partially loaded From: Tim Chen To: Peter Zijlstra Cc: Ingo Molnar , Andi Kleen , Michel Lespinasse , Rik van Riel , Peter Hurley , Jason Low , Davidlohr Bueson , linux-kernel@vger.kernel.org In-Reply-To: <20140615161921.GL11371@laptop.programming.kicks-ass.net> References: <1402608359.2970.548.camel@schen9-DESK> <20140615161921.GL11371@laptop.programming.kicks-ass.net> Content-Type: text/plain; charset="UTF-8" Date: Mon, 16 Jun 2014 08:50:07 -0700 Message-ID: <1402933807.2970.577.camel@schen9-DESK> Mime-Version: 1.0 X-Mailer: Evolution 2.32.3 (2.32.3-1.fc14) Content-Transfer-Encoding: 7bit Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On Sun, 2014-06-15 at 18:19 +0200, Peter Zijlstra wrote: > On Thu, Jun 12, 2014 at 02:25:59PM -0700, Tim Chen wrote: > > @@ -2630,7 +2630,7 @@ static inline struct task_struct * > > pick_next_task(struct rq *rq, struct task_struct *prev) > > { > > const struct sched_class *class = &fair_sched_class; > > - struct task_struct *p; > > + struct task_struct *p = NULL; > > > > /* > > * Optimization: we know that if all tasks are in > > @@ -2638,9 +2638,13 @@ pick_next_task(struct rq *rq, struct task_struct *prev) > > */ > > if (likely(prev->sched_class == class && > > rq->nr_running == rq->cfs.h_nr_running)) { > > - p = fair_sched_class.pick_next_task(rq, prev); > > - if (unlikely(p == RETRY_TASK)) > > - goto again; > > + > > + /* If no cpu has more than 1 task, skip */ > > + if (rq->nr_running > 0 || rq->rd->overload) { > > + p = fair_sched_class.pick_next_task(rq, prev); > > + if (unlikely(p == RETRY_TASK)) > > + goto again; > > + } > > > > /* assumes fair_sched_class->next == idle_sched_class */ > > if (unlikely(!p)) > > > Please move this into pick_next_task_fair(). You're slowing down the > important fast path of picking a task when there actually is something > to do. Will do. > > Also, its a layering violation -- the idle balance things you're trying > to avoid is a fair_sched_class affair. > > > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c > > index 9855e87..00ab38c 100644 > > --- a/kernel/sched/fair.c > > +++ b/kernel/sched/fair.c > > @@ -5863,7 +5863,8 @@ static inline int sg_capacity(struct lb_env *env, struct sched_group *group) > > */ > > static inline void update_sg_lb_stats(struct lb_env *env, > > struct sched_group *group, int load_idx, > > - int local_group, struct sg_lb_stats *sgs) > > + int local_group, struct sg_lb_stats *sgs, > > + bool *overload) > > { > > unsigned long load; > > int i; > > @@ -5881,6 +5882,8 @@ static inline void update_sg_lb_stats(struct lb_env *env, > > > > sgs->group_load += load; > > sgs->sum_nr_running += rq->nr_running; > > + if (overload && rq->nr_running > 1) > > + *overload = true; > > #ifdef CONFIG_NUMA_BALANCING > > sgs->nr_numa_running += rq->nr_numa_running; > > sgs->nr_preferred_running += rq->nr_preferred_running; > > @@ -5991,6 +5994,7 @@ static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sd > > struct sched_group *sg = env->sd->groups; > > struct sg_lb_stats tmp_sgs; > > int load_idx, prefer_sibling = 0; > > + bool overload = false; > > > > if (child && child->flags & SD_PREFER_SIBLING) > > prefer_sibling = 1; > > @@ -6011,7 +6015,13 @@ static inline void update_sd_lb_stats(struct lb_env *env, struct sd_lb_stats *sd > > update_group_power(env->sd, env->dst_cpu); > > } > > > > - update_sg_lb_stats(env, sg, load_idx, local_group, sgs); > > + if (env->sd->parent) > > + update_sg_lb_stats(env, sg, load_idx, local_group, sgs, > > + NULL); > > + else > > + /* gather overload info if we are at root domain */ > > + update_sg_lb_stats(env, sg, load_idx, local_group, sgs, > > + &overload); > > > > if (local_group) > > goto next_group; > > @@ -6045,6 +6055,15 @@ next_group: > > > > if (env->sd->flags & SD_NUMA) > > env->fbq_type = fbq_classify_group(&sds->busiest_stat); > > + > > + if (!env->sd->parent) { > > + /* update overload indicator if we are at root domain */ > > + int i = cpumask_first(sched_domain_span(env->sd)); > > + struct rq *rq = cpu_rq(i); > > + if (rq->rd->overload != overload) > > + rq->rd->overload = overload; > > + } > > + > > } > > > > /** > > The worry I have is that this update is 'slow'. We could have grown many > tasks since the last update. The update to turn on the indicator is immediate and triggered in add_nr_running. So if there are more than one tasks on any cpu, we start load balancing again right away. It is only the clearing of the indicator in update_sd_lb_stats that takes time. That does no harm as the cleared indicator is for the skipping of load balance, which can be delayed. Thanks. Tim