Re: [RFC] scheduler: improve SMP fairness in CFS

[Ingo Molnar <mingo@elte.hu>]

* Tong Li <tong.n.li@intel.com> wrote:

> Thanks for the patch. It doesn't work well on my 8-way box. Here's the 
> output at two different times. It's also changing all the time.

you need to measure it over longer periods of time. Its not worth 
balancing for such a thing in any high-frequency manner. (we'd trash the 
cache constantly migrating tasks back and forth.)

> [time 2]
>  3721 tongli    20   0  7864  108   16 R   91  0.0   1:44.67 loop
>  3720 tongli    20   0  7864  108   16 R   89  0.0   1:40.55 loop
>  3722 tongli    20   0  7864  108   16 R   86  0.0   1:41.40 loop
>  3723 tongli    20   0  7864  108   16 R   82  0.0   1:38.94 loop
>  3715 tongli    20   0  7864  108   16 R   82  0.0   1:43.09 loop
>  3717 tongli    20   0  7864  108   16 R   79  0.0   1:46.38 loop
>  3718 tongli    20   0  7864  108   16 R   79  0.0   1:40.72 loop
>  3714 tongli    20   0  7864  108   16 R   77  0.0   1:40.17 loop
>  3719 tongli    20   0  7864  108   16 R   73  0.0   1:35.41 loop
>  3716 tongli    20   0  7864  108   16 R   61  0.0   1:38.66 loop

the runtime ranges from 1:38.66 to 1:46.38 - that's a +-3% spread which 
is pretty acceptable for ~90 seconds of runtime. Note that the 
percentages you see above were likely done over a shorter period that's 
why you see them range from 61% to 91%.

> I'm running 2.6.23-rc1 with your patch on a two-socket quad-core 
> system. Top refresh rate is the default 3s.

the 3s is the problem: change that to 60s! We no way want to 
over-migrate for SMP fairness, the change i did gives us reasonable 
long-term SMP fairness without the need for high-rate rebalancing.

	Ingo
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/