Re: [PATCH 2/4] sched,numa: weigh nearby nodes for task placement on complex NUMA topologies

[Rik van Riel <riel@redhat.com>]

On 05/09/2014 06:11 AM, Peter Zijlstra wrote:
> On Thu, May 08, 2014 at 01:23:29PM -0400, riel@redhat.com wrote:
>> @@ -930,7 +987,7 @@ static inline unsigned long group_faults_cpu(struct numa_group *group, int nid)
>>    */
>>   static inline unsigned long task_weight(struct task_struct *p, int nid)
>>   {
>> -	unsigned long total_faults;
>> +	unsigned long total_faults, score;
>>
>>   	if (!p->numa_faults_memory)
>>   		return 0;
>> @@ -940,15 +997,32 @@ static inline unsigned long task_weight(struct task_struct *p, int nid)
>>   	if (!total_faults)
>>   		return 0;
>>
>> -	return 1000 * task_faults(p, nid) / total_faults;
>> +	score = 1000 * task_faults(p, nid);
>> +	score += nearby_nodes_score(p, nid, true);
>> +
>> +	score /= total_faults;
>> +
>> +	return score;
>>   }
>>
>>   static inline unsigned long group_weight(struct task_struct *p, int nid)
>>   {
>> -	if (!p->numa_group || !p->numa_group->total_faults)
>> +	unsigned long total_faults, score;
>> +
>> +	if (!p->numa_group)
>> +		return 0;
>> +
>> +	total_faults = p->numa_group->total_faults;
>> +
>> +	if (!total_faults)
>>   		return 0;
>>
>> -	return 1000 * group_faults(p, nid) / p->numa_group->total_faults;
>> +	score = 1000 * group_faults(p, nid);
>> +	score += nearby_nodes_score(p, nid, false);
>> +
>> +	score /= total_faults;
>> +
>> +	return score;
>>   }
>
> OK, and that's just sad..
>
> See task_numa_placement(), which does:
>
> 	for_each_online_node(nid) {
> 		weight = task_weight(p, nid) + group_weight(p, nid);
> 		if (weight > max_weight) {
> 			max_weight = weight;
> 			max_nid = nid;
> 		}
> 	}
>
> So not only is that loop now O(nr_nodes^2), the inner loops doubly
> iterates all nodes.

I am not too worried about task_numa_placement, but you are
right that this may well be much too expensive for more
frequently called code like migrate_improves_locality.

Having said that, grouping related tasks together on nearby
nodes does seem to bring significant performance gains.

Do you have any ideas on other ways we can achieve that
grouping?

> Also, {task,group}_weight() functions were like cheap-ish (/me mumbles
> something about people using !2^n scaling factors for no sane reason).
> And they're used all over with that in mind.
>
> But look what you did to migrate_improves_locality(), that will now
> iterate all nodes _4_ times, and its called for every single task we try
> and migrate during load balance, while holding rq->lock.
>
>