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

[Paul Turner]

On Mon, May 6, 2013 at 3:19 AM, Peter Zijlstra <a.p.zijlstra@xxxxxxxxx> wrote:
> On Mon, May 06, 2013 at 01:46:19AM -0700, Paul Turner wrote:
>> On Sun, May 5, 2013 at 6:45 PM, Alex Shi <alex.shi@xxxxxxxxx> wrote:
>> > @@ -2536,7 +2536,7 @@ static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
>> >  void update_idle_cpu_load(struct rq *this_rq)
>> >  {
>> >         unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
>> > -       unsigned long load = this_rq->load.weight;
>> > +       unsigned long load = (unsigned long)this_rq->cfs.runnable_load_avg;
>>
>> We should be minimizing:
>>   Variance[ for all i ]{ cfs_rq[i]->runnable_load_avg +
>> cfs_rq[i]->blocked_load_avg }
>>
>> 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;
>>
>> 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.
>
> Ah indeed. I overlooked that.
>
>> 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.
>
> So I suspect (haven't gone back in history to verify) that load_idx mostly
> comes from the fact that our balance passes happen more and more slowly the
> bigger the domains get.
>
> In that respect it makes sense to equate load_idx to sched_domain::level;
> higher domains balance slower and would thus want a longer-term average to base
> decisions on.
>
> So what we would want is means to get sane longer term averages.

Yeah, most of the rationale is super hand-wavy; especially the fairly
arbitrary choice of periods (e.g. busy_idx vs newidle).

I think the other rationale is:
  For smaller indicies (e.g. newidle) we speed up response time by
also cutting motion out of the averages.

The runnable_avgs themselves actually have a fair bit of history in
them already (50% is last 32ms); but given that they don't need to be
cut-off to respond to load being migrated I'm guessing we could
actually potentially get by with just "instaneous" and "use averages"
where appropriate?

We always end up having to re-pick/tune them based on a variety of
workloads; if we can eliminate them I think it would be a win.
--
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/