Re: [RFC tg_shares_up improvements - v1 00/12] [RFC tg_shares_up - v100/12] Reducing cost of tg->shares distribution

[Paul Turner]

On Wed, Oct 20, 2010 at 11:36 PM, Paul Turner <pjt@xxxxxxxxxx> wrote:
> On Sat, Oct 16, 2010 at 12:46 PM, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>> On Fri, 2010-10-15 at 21:43 -0700, pjt@xxxxxxxxxx wrote:
>>> Hi all,
>>>
>>> Peter previously posted a patchset that attempted to improve the problem of
>>> task_group share distribution.  This is something that has been a long-time
>>> pain point for group scheduling.  The existing algorithm considers
>>> distributions on a per-cpu-per-domain basis and carries a fairly high update
>>> overhead, especially on larger machines.
>>>
>>> I was previously looking at improving this using Fenwick trees to allow a
>>> single sum without the exorbitant cost but then Peter's idea above was better :).
>>>
>>> The kernel is that by monitoring the average contribution to load on a
>>> per-cpu-per-taskgroup basis we can distribute the weight for which we are
>>> expected to consume.
>>>
>>> This set extends the original posting with a focus on increased fairness and
>>> reduced convergence (to true average) time.  In particular the case of large
>>> over-commit in the case of a distributed wake-up is a concern which is now
>>> fairly well addressed.
>>>
>>> Obviously everything's experimental but it should be stable/fair.
>>
>> I like what you've done with it, my only worry is 10/12 where you allow
>> for extra updates to the global state -- I think they should be fairly
>> limited in number, and I can see the need for the update if we get too
>> far out of whack, but it is something to look at while testing this
>> stuff.
>>
>
> So my original answer here was to only update when there was load and
> it was > n% delta which stops 1 thread waking up and sleeping from
> thrashing it, but the 2 thread case is just as obviously broken for
> any n.  It needs a rate limit but I'm sort of loathe to introduce
> _another_ set of timestamps.  I don't suppose there's much harm in
> doing so though and I don't think it's going to be clean to overload
> one of the existing ones so perhaps another counter is the answer.
>
> I'll make sure this is addressed in v2.

Ahh -- I remember my original reasoning here now:

These global updates are rate limited since they only occur when we
fold the averaging period thus are limited to occur at a rate of
period/2 (e.g. each time we fold).

One concern is we could get a local "storm" since several cgroups
could trip this limit at the same time.. but it's also obviously not
going to be any worse than an update_shares_cpu().

While there may be a concern on *giant* cpu systems, I think we
already have that problem in both the original and (to a lesser
extent) current approaches.

>
>>> TODO:
>>> - Validate any RT interaction
>>
>> I don't think there's anything to worry about there, the only
>> interaction which there is between this and the rt scheduling classes is
>> the initial sharing of the load-avg window, but you 'cure' that in 7/12.
>>
>> (I think that sysctl wants a _us postfix someplace and we thus want some
>> NSEC_PER_USEC multiplication in there).
>>
>
> Yes -- updated, thanks.
>
>>> - Continue collecting/analyzing performance and fairness data
>>
>> Yes please ;-), I'll try and run this on some machines as well.
>>
>>> - Should the shares period just be the sched_latency?
>>
>> Interesting idea.. lets keep it a separate sysctl for now for easy
>> tuning, if things settle down and we're still good in that range we can
>> consider merging them.
>>
>>
>>
>
--
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/