Re: [PATCH] sched/numa: advanced per-cgroup numa statistic

[王贇 <yun.wang@linux.alibaba.com>]

On 2019/10/28 下午9:02, Peter Zijlstra wrote:
[snip]
>> +	tg = task_group(p);
>> +	while (tg) {
>> +		/* skip account when there are no faults records */
>> +		if (idx != -1)
>> +			this_cpu_inc(tg->numa_stat->locality[idx]);
>> +
>> +		this_cpu_inc(tg->numa_stat->jiffies);
>> +
>> +		tg = tg->parent;
>> +	}
>> +
>> +	rcu_read_unlock();
>> +}
> 
> Thing is, we already have a cgroup hierarchy walk in the tick; see
> task_tick_fair().
> 
> On top of that, you're walking an entirely different set of pointers,
> instead of cfs_rq, you're walking tg->parent, which pretty much
> guarantees you're adding even more cache misses.
> 
> How about you stick those numa_stats in cfs_rq (with cacheline
> alignment) and see if you can frob your update loop into the cgroup walk
> we already do.

Thanks for the reply :-)

The hierarchy walk here you mean is the loop of entity_tick(), correct?

Should working if we introduce the per-cfs_rq numa_stat accounting and
do update there, I'll try to reform the implementation in next version.

Regards,
Michael Wang

>