Re: [patch v3 0/8] sched: use runnable avg in load balance

[Michael Wang <wangyun@linux.vnet.ibm.com>]

On 04/03/2013 12:28 PM, Alex Shi wrote:
[snip]
> 
> but the patch may cause some unfairness if this/prev cpu are not burst at 
> same time. So could like try the following patch?

I will try it later, some doubt below :)

[snip]
> +
> +	if (cpu_rq(this_cpu)->avg_idle < sysctl_sched_migration_cost ||
> +		cpu_rq(prev_cpu)->avg_idle < sysctl_sched_migration_cost)
> +		burst= 1;
> +
> +	/* use instant load for bursty waking up */
> +	if (!burst) {
> +		load = source_load(prev_cpu, idx);
> +		this_load = target_load(this_cpu, idx);
> +	} else {
> +		load = cpu_rq(prev_cpu)->load.weight;
> +		this_load = cpu_rq(this_cpu)->load.weight;

Ok, my understanding is, we want pull if 'prev_cpu' is burst, and don't
want pull if 'this_cpu' is burst, correct?

And we do this by guess the load higher or lower, is that right?

And I think target_load() is capable enough to choose the higher load,
if 'cpu_rq(cpu)->load.weight' is really higher, the results will be the
same.

So what about this:

	/* prefer higher load if burst */
	load = burst_prev ?
		target_load(prev_cpu, idx) : source_load(prev_cpu, idx);

	this_load = target_load(this_cpu, idx);

Regards,
Michael Wang

> +	}
> 
>  	/*
>  	 * If sync wakeup then subtract the (maximum possible)
>