[RFC PATCH v3 3/6] sched: pack small tasks

[Vincent Guittot]

On 26 March 2013 13:46, Peter Zijlstra <peterz at infradead.org> wrote:
> On Fri, 2013-03-22 at 13:25 +0100, Vincent Guittot wrote:
>> During the creation of sched_domain, we define a pack buddy CPU for
>> each CPU
>> when one is available. We want to pack at all levels where a group of
>> CPU can
>> be power gated independently from others.
>> On a system that can't power gate a group of CPUs independently, the
>> flag is
>> set at all sched_domain level and the buddy is set to -1. This is the
>> default
>> behavior.
>> On a dual clusters / dual cores system which can power gate each core
>> and
>> cluster independently, the buddy configuration will be :
>>
>>       | Cluster 0   | Cluster 1   |
>>       | CPU0 | CPU1 | CPU2 | CPU3 |
>> -----------------------------------
>> buddy | CPU0 | CPU0 | CPU0 | CPU2 |
>
> I suppose this is adequate for the 'small' systems you currently have;
> but given that Samsung is already bragging with its 'octo'-core Exynos
> 5 (4+4 big-little thing) does this solution scale?

The packing is only done at MC and CPU level to minimize the number of
transition.

>
> Isn't this basically related to picking the NO_HZ cpu; if the system
> isn't fully symmetric with its power gates you want the NO_HZ cpu to be
> the 'special' cpu. If it is symmetric we really don't care which core
> is left 'running' and we can even select a new pack cpu from the idle
> cores once the old one is fully utilized.

I agree that on a symmetric system, we don't really care about which
core is selected but we want to use the same one whenever possible to
prevent a ping pong between several cores or groups of cores, which is
power consuming. By forcing a NOHZ cpu, your background activity will
smoothly pack on this CPU and will not be spread on your system.
When a CPU is fully loaded, we don't fall in a low CPU load use case
and the periodic load balance can handle the situation to select a new
target CPU which is close to the buddy CPU

>
> Re-using (or integrating) with NO_HZ has the dual advantage that you'll
> make NO_HZ do the right thing for big-little (you typically want a
> little core to be the one staying 'awake' and once someone makes NO_HZ
> scale this all gets to scale along with it.
>

I think that you have answered to this question in your comment of
patch 5, isn't it?

Vincent