Re: [PATCH 0/2] /proc/stat: Reduce irqs counting performance overhead

[Waiman Long <longman@redhat.com>]

On 01/08/2019 11:11 AM, Michal Hocko wrote:
> On Tue 08-01-19 13:04:22, Dave Chinner wrote:
>> On Mon, Jan 07, 2019 at 05:41:39PM -0500, Waiman Long wrote:
>>> On 01/07/2019 05:32 PM, Dave Chinner wrote:
>>>> On Mon, Jan 07, 2019 at 10:12:56AM -0500, Waiman Long wrote:
>>>>> As newer systems have more and more IRQs and CPUs available in their
>>>>> system, the performance of reading /proc/stat frequently is getting
>>>>> worse and worse.
>>>> Because the "roll-your-own" per-cpu counter implementaiton has been
>>>> optimised for low possible addition overhead on the premise that
>>>> summing the counters is rare and isn't a performance issue. This
>>>> patchset is a direct indication that this "summing is rare and can
>>>> be slow" premise is now invalid.
>>>>
>>>> We have percpu counter infrastructure that trades off a small amount
>>>> of addition overhead for zero-cost reading of the counter value.
>>>> i.e. why not just convert this whole mess to percpu_counters and
>>>> then just use percpu_counter_read_positive()? Then we just don't
>>>> care how often userspace reads the /proc file because there is no
>>>> summing involved at all...
>>>>
>>>> Cheers,
>>>>
>>>> Dave.
>>> Yes, percpu_counter_read_positive() is cheap. However, you still need to
>>> pay the price somewhere. In the case of percpu_counter, the update is
>>> more expensive.
>> Ummm, that's exactly what I just said. It's a percpu counter that
>> solves the "sum is expensive and frequent" problem, just like you
>> are encountering here. I do not need basic scalability algorithms
>> explained to me.
>>
>>> I would say the percentage of applications that will hit this problem is
>>> small. But for them, this problem has some significant performance overhead.
>> Well, duh!
>>
>> What I was suggesting is that you change the per-cpu counter
>> implementation to the /generic infrastructure/ that solves this
>> problem, and then determine if the extra update overhead is at all
>> measurable. If you can't measure any difference in update overhead,
>> then slapping complexity on the existing counter to attempt to
>> mitigate the summing overhead is the wrong solution.
>>
>> Indeed, it may be that you need o use a custom batch scaling curve
>> for the generic per-cpu coutner infrastructure to mitigate the
>> update overhead, but the fact is we already have generic
>> infrastructure that solves your problem and so the solution should
>> be "use the generic infrastructure" until it can be proven not to
>> work.
>>
>> i.e. prove the generic infrastructure is not fit for purpose and
>> cannot be improved sufficiently to work for this use case before
>> implementing a complex, one-off snowflake counter implementation...
> Completely agreed! Apart from that I find that conversion to a generic
> infrastructure worth even if that doesn't solve the problem at hands
> completely. If for no other reasons then the sheer code removal as kstat
> is not really used for anything apart from this accounting AFAIR. The
> less ad-hoc code we have the better IMHO.
>
> And to the underlying problem. Some proc files do not scale on large
> machines. Maybe it is time to explain that to application writers that
> if they are collecting data too agressively then it won't scale. We can
> only do this much. Lying about numbers by hiding updates is, well,
> lying and won't solve the underlying problem. 

I would not say it is lying. As I said in the changelog, reading
/proc/stat infrequently will give the right counts. Only when it is read
frequently that the data may not be up-to-date. Using
percpu_counter_sum_positive() as suggested by Dave will guarantee that
the counts will likely be off by a certain amount too. So it is also a
trade-off between accuracy and performance.

Cheers,
Longman