Re: [PATCH v5 5/7] sched: compute runnable load avg in cpu_load and cpu_avg_load_per_task

From: Paul Turner
Date: Mon May 06 2013 - 14:35:04 EST


On Mon, May 6, 2013 at 8:00 AM, Alex Shi <alex.shi@xxxxxxxxx> wrote:
>>
>> blocked_load_avg is the expected "to wake" contribution from tasks
>> already assigned to this rq.
>>
>> e.g. this could be:
>> load = this_rq->cfs.runnable_load_avg + this_rq->cfs.blocked_load_avg;
>
> Current load balance doesn't consider slept task's load which is
> represented by blocked_load_avg. And the slept task is not on_rq, so
> consider it in load balance is a little strange.

The load-balancer has a longer time horizon; think of blocked_loag_avg
to be a signal for the load, already assigned to this cpu, which is
expected to appear (within roughly the next quantum).

Consider the following scenario:

tasks: A,B (40% busy), C (90% busy)

Suppose we have:
CPU 0: CPU 1:
A C
B

Then, when C blocks the load balancer ticks.

If we considered only runnable_load then A or B would be eligible for
migration to CPU 1, which is essentially where we are today.

>
> But your concern is worth to try. I will change the patchset and give
> the testing results.
>
>>
>> Although, in general I have a major concern with the current implementation:
>>
>> The entire reason for stability with the bottom up averages is that
>> when load migrates between cpus we are able to migrate it between the
>> tracked sums.
>>
>> Stuffing observed averages of these into the load_idxs loses that
>> mobility; we will have to stall (as we do today for idx > 0) before we
>> can recognize that a cpu's load has truly left it; this is a very
>> similar problem to the need to stably track this for group shares
>> computation.
>>
>> To that end, I would rather see the load_idx disappear completely:
>> (a) We can calculate the imbalance purely from delta (runnable_avg +
>> blocked_avg)
>> (b) It eliminates a bad tunable.
>
> I also show the similar concern of load_idx months ago. seems overlooked. :)
>>
>>> - return cpu_rq(cpu)->load.weight;
>>> + return (unsigned long)cpu_rq(cpu)->cfs.runnable_load_avg;
>>
>> Isn't this going to truncate on the 32-bit case?
>
> I guess not, the old load.weight is unsigned long, and runnable_load_avg
> is smaller than the load.weight. so it should be fine.
>
> btw, according to above reason, guess move runnable_load_avg to
> 'unsigned long' type is ok, do you think so?
>

Hmm, so long as it's unsigned long and not u32 that should be OK.

>From a technical standpoint:
We make the argument that we run out of address space before we can
overflow load.weight in the 32-bit case, we can make the same argument
here.

>
> --
> Thanks
> Alex
--
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/