Re: [PATCH] sched/fair: update scale invariance of pelt

[Vincent Guittot]

On 25 November 2015 at 10:24, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> On Tue, Nov 24, 2015 at 02:49:30PM +0100, Vincent Guittot wrote:
>> Instead of scaling the complete value of PELT algo, we should only scale
>> the running time by the current capacity of the CPU. It seems more correct
>> to only scale the running time because the non running time of a task
>> (sleeping or waiting for a runqueue) is the same whatever the current freq
>> and the compute capacity of the CPU.
>
> So I'm leaning towards liking this; however with your previous example
> of 3 cpus and 7 tasks, where CPU0-1 are 'little' and of half the
> capacity as the 'big' CPU2, with 2 tasks on CPU0-1 each and 3 tasks on
> CPU2.
>
> This would result, for CPU0, in a load of 100% wait time + 100% runtime,
> scaling the runtime 50% will get you a total load of 150%.
>
> For CPU2 we get 100% runtime and 200% wait time, no scaling, for a total
> load of 300%.
>
> So the CPU0-1 cluster has a 300% load and the CPU2 'cluster' has a 300%
> load, even though the actual load is not actually equal, CPUs0-1
> combined have the same capacity as CPU2, so it should be 4-4 tasks for
> an equal balance.

With the example above, we have (after that everything has reached
their stable value)
With the mainline:
load_avg of CPU0 : 2048 and load_avg of each task should be 1024
load_avg of CPU1 : 2048 and load_avg of each task should be 1024
load_avg of CPU2 : 3072 and load_avg of each task should be 1024

With this patch which now includes the cpu invariance in the
calculation of load_avg
load_avg of CPU0 : 2048 and load_avg of each task should be 1024
load_avg of CPU1 : 2048 and load_avg of each task should be 1024
load_avg of CPU2 : 3072 and load_avg of each task should be 1024

The main difference will be in the time needed to reach these values.
CPU2 will reach 95% of the final value in 136ms whereas the load_avg
of CPU0 and CPU1 should be around 789 at that time and will reach the
same value than CPU2 after additional 136ms

Regards,
Vincent

>
>
> So I'm not sure the claim of comparable between CPUs stands. Still it is
> an interesting idea and I will consider it more.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/