On 9 October 2014 17:30, Peter Zijlstra peterz@infradead.org wrote:
On Thu, Oct 09, 2014 at 04:59:36PM +0200, Vincent Guittot wrote:
On 9 October 2014 13:23, Peter Zijlstra peterz@infradead.org wrote:
On Tue, Oct 07, 2014 at 02:13:32PM +0200, Vincent Guittot wrote:
+++ b/kernel/sched/fair.c @@ -5896,6 +5896,18 @@ fix_small_capacity(struct sched_domain *sd, struct sched_group *group) }
/*
- Check whether the capacity of the rq has been noticeably reduced by side
- activity. The imbalance_pct is used for the threshold.
- Return true is the capacity is reduced
- */
+static inline int +check_cpu_capacity(struct rq *rq, struct sched_domain *sd) +{
return ((rq->cpu_capacity * sd->imbalance_pct) <(rq->cpu_capacity_orig * 100));+}
+/*
- Group imbalance indicates (and tries to solve) the problem where balancing
- groups is inadequate due to tsk_cpus_allowed() constraints.
@@ -6567,6 +6579,14 @@ static int need_active_balance(struct lb_env *env) */ if ((sd->flags & SD_ASYM_PACKING) && env->src_cpu > env->dst_cpu) return 1;
/** The src_cpu's capacity is reduced because of other* sched_class or IRQs, we trig an active balance to move the* task*/if (check_cpu_capacity(env->src_rq, sd))return 1; }So does it make sense to first check if there's a better candidate at all? By this time we've already iterated the current SD while trying regular load balancing, so we could know this.
i'm not sure to completely catch your point. Normally, f_b_g and f_b_q have already looked at the best candidate when we call need_active_balance. And src_cpu has been elected. Or i have missed your point ?
Yep you did indeed miss my point.
So I've always disliked this patch for its arbitrary nature, why unconditionally try and active balance every time there is 'some' RT/IRQ usage, it could be all CPUs are over that arbitrary threshold and we'll end up active balancing for no point.
So, since we've already iterated all CPUs in our domain back in update_sd_lb_stats() we could have computed the CFS fraction:
1024 * capacity / capacity_origfor every cpu and collected the min/max of this. Then we can compute if src is significantly (and there I suppose we can indeed use imb) affected compared to others.
ok, so we should put additional check in f_b_g to be sure that we will jump to force_balance only if there will be a real gain by moving the task on the local group (from an available capacity for the task point of view) and probably in f_b_q too