Re: hackbench regression with kernel 2.6.32-rc1

["Zhang, Yanmin" <yanmin_zhang@linux.intel.com>]

On Fri, 2009-10-16 at 13:06 +0200, Mike Galbraith wrote:
> On Tue, 2009-10-13 at 11:12 +0800, Zhang, Yanmin wrote:
> 
> > NEXT_BUDDY has no help on volanoMark and tbench.
> 
> Can you try the patch below please?  It does tries to preserve buddy
> affinity where possible, and mitigates over-preemption by strengthening
> buddies a bit.  It improves vmark here by ~7%.
I ran some benchmarks against 2.6.32-rc1+Peter_2_patches+below_patch.
Below result is against 2.6.32-rc1.
hackbench result has about 10% improvement on stoakley (2*4 cores) and
tigerton (4*4 cores).
tbench still has about 5% regression on stoakley and tigerton.
VolanoMark has 33% regression on tigerton, but has 2% improvement on stoakley.

I also ran the benchmarks against the latest tips/master and got the similiar
results like above testing.

The testing against tips on Nehalem machine didn't show much improvement/regression.

> 
> diff --git a/kernel/sched.c b/kernel/sched.c
> index 00f9e71..fb025d4 100644
> --- a/kernel/sched.c
> +++ b/kernel/sched.c
> @@ -2007,8 +2007,12 @@ task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
>  
>  	/*
>  	 * Buddy candidates are cache hot:
> +	 *
> +	 * Do not honor buddies if there may be nothing else to
> +	 * prevent us from becoming idle.
>  	 */
>  	if (sched_feat(CACHE_HOT_BUDDY) &&
> +			task_rq(p)->nr_running >= sched_nr_latency &&
>  			(&p->se == cfs_rq_of(&p->se)->next ||
>  			 &p->se == cfs_rq_of(&p->se)->last))
>  		return 1;
> diff --git a/kernel/sched_fair.c b/kernel/sched_fair.c
> index c32c3e6..428bf55 100644
> --- a/kernel/sched_fair.c
> +++ b/kernel/sched_fair.c
> @@ -863,18 +863,20 @@ static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
>  	struct sched_entity *se = __pick_next_entity(cfs_rq);
>  	struct sched_entity *buddy;
>  
> -	if (cfs_rq->next) {
> +	if (cfs_rq->next && sched_feat(NEXT_BUDDY)) {
>  		buddy = cfs_rq->next;
> -		cfs_rq->next = NULL;
> -		if (wakeup_preempt_entity(buddy, se) < 1)
> +		if (wakeup_preempt_entity(buddy, se) < 1) {
> +			cfs_rq->next = NULL;
>  			return buddy;
> +		}
>  	}
>  
> -	if (cfs_rq->last) {
> +	if (cfs_rq->last && sched_feat(LAST_BUDDY)) {
>  		buddy = cfs_rq->last;
> -		cfs_rq->last = NULL;
> -		if (wakeup_preempt_entity(buddy, se) < 1)
> +		if (wakeup_preempt_entity(buddy, se) < 1) {
> +			cfs_rq->last = NULL;
>  			return buddy;
> +		}
>  	}
>  
>  	return se;
> @@ -1600,9 +1602,9 @@ static void check_preempt_wakeup(struct rq *rq, struct task_struct *p, int wake_
>  	 * Also, during early boot the idle thread is in the fair class, for
>  	 * obvious reasons its a bad idea to schedule back to the idle thread.
>  	 */
> -	if (sched_feat(LAST_BUDDY) && likely(se->on_rq && curr != rq->idle))
> +	if (!(wake_flags & WF_FORK) && likely(se->on_rq && curr != rq->idle))
>  		set_last_buddy(se);
> -	if (sched_feat(NEXT_BUDDY) && !(wake_flags & WF_FORK))
> +	if (!(wake_flags & WF_FORK))
>  		set_next_buddy(pse);
>  
>  	/*
> 
>