Re: [patch v8 6/9] sched: compute runnable load avg in cpu_load and cpu_avg_load_per_task

[Alex Shi <alex.shi@intel.com>]

On 06/19/2013 04:15 PM, Alex Shi wrote:
> On 06/18/2013 05:44 PM, Alex Shi wrote:
>>
>>>
>>> Paul, could I summary your point here:
>>> keep current weighted_cpu_load, but add blocked load avg in
>>> get_rq_runnable_load?
>>>
>>> I will test this change.
>>
>> Current testing(kbuild, oltp, aim7) don't show clear different on my NHM EP box
>> between the following and the origin patch, 
>> the only different is get_rq_runnable_load added blocked_load_avg. in SMP
>> will test more cases and more box.
> 
> I tested the tip/sched/core, tip/sched/core with old patchset and
> tip/schec/core with the blocked_load_avg on Core2 2S, NHM EP, IVB EP,
> SNB EP 2S and SNB EP 4S box, with benchmark kbuild, sysbench oltp,
> hackbench, tbench, dbench.
> 
> blocked_load_avg VS origin patchset, oltp has suspicious 5% and
> hackbench has 3% drop on NHM EX; dbench has suspicious 6% drop on NHM
> EP. other benchmarks has no clear change on all other machines.
> 
> origin patchset VS sched/core, hackbench rise 20% on NHM EX, 60% on SNB
> EP 4S, and 30% on IVB EP. others no clear changes.
> 

>> +#ifdef CONFIG_SMP
>> +unsigned long get_rq_runnable_load(struct rq *rq)
>> +{
>> +	return rq->cfs.runnable_load_avg + rq->cfs.blocked_load_avg;

According to above testing result, with blocked_load_avg is still a
slight worse than without it.

when the blocked_load_avg added here, it will impact nohz idle balance
and periodic balance in update_sg_lb_stats() when the idx is not 0.
As to nohz idle balance, blocked_load_avg should be too small to have
big effect.
As to in update_sg_lb_stats(), since it only works when _idx is not 0,
that means the blocked_load_avg was decay again in update_cpu_load. That
reduce its impact.

So, could I say, at least in above testing, blocked_load_avg should be
keep away from balance?


-- 
Thanks
    Alex