Re: [RFC PATCH v3 0/3] sched: simplify the select_task_rq_fair()

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

On 02/22/2013 05:57 PM, Peter Zijlstra wrote:
> On Fri, 2013-02-22 at 17:11 +0800, Michael Wang wrote:
> 
>> Ok, it do looks like wake_affine() lost it's value...
> 
> I'm not sure we can say that on this one benchmark, there's a
> preemption advantage to running on a single cpu for pipe-test as well.
> We'd need to create a better benchmark to test this, one that has some
> actual data payload and control over the initial spread of the tasks or
> so.
> 
>>> Now as far as I can see there's two options, either we find there's
>>> absolutely no benefit in wake_affine() as it stands today and we simply
>>> disable/remove it, or we go fix it. What we don't do is completely
>>> wreck it at atrocious cost.
>>
>> I get your point, we should replace wake_affine() with some feature
>> which could really achieve the goal to make client and server on same cpu.
>>
>> But is the logical that the waker/wakee are server/client(or reversed)
>> still works now? that sounds a little arbitrary to me...
> 
> Ah, its never really been about server/client per-se. Its just a
> specific example -- one that breaks down with the 1:n pgbench
> situation.
> 
> Wakeups in general can be considered to be a relation, suppose a
> hardware interrupt that received some data from a device and issues a
> wakeup to a task to consume this data. What CPU would be better suited
> to process this data then the one where its already cache hot.

I see, honestly, I realized that I have underestimated the benefit we
gain from it when saw your testing results...

We do need some better approach to replace wake_affine(), hmm...I need a
draft board now...

Regards,
Michael Wang

>