Re: [RFC PATCH 0/3] sched: Introduce Window Assisted Load Tracking

[Vikram Mulukutla]

On 2016-10-28 01:49, Peter Zijlstra wrote:
> On Fri, Oct 28, 2016 at 12:57:05AM -0700, Vikram Mulukutla wrote:
> > On 2016-10-28 00:49, Peter Zijlstra wrote:
> > >On Fri, Oct 28, 2016 at 12:10:39AM -0700, Vikram Mulukutla wrote:
> > >>This RFC patch has been tested on live X86 machines with the following
> > >>sanity
> > >>and benchmark results (thanks to Juri Lelli, Dietmar Eggeman, Patrick
> > >>Bellasi
> > >>for initial code reviews):
> > >>
> > >>(Tested on an Intel i7 2nd generation CPU, 8GB RAM, Nvidia GTX950Ti
> > >>graphics,
> > >>with the same frequency list as above. Running Ubuntu 16.04 on a v4.8.2
> > >>baseline. WALT window size was 10ms. Only deltas above 3% are considered
> > >>non-noise.Power measured with Intel RAPL counters)
> > >
> > >Was this comparison done using the use_walt_metric sysctl knob?
> >
> > Yes, it was. You will want to see numbers against a pure 4.8.2 without 
> > any
> > of the WALT code, correct?
>
> Yep, because with the sysctl we still run all the accounting code. So
> esp things like the hackbench run are meaningless (note that even the
> CONFIG thing doesn't take out everything).
>
> Also, I think it makes sense to always (also) compare against the
> "performance" governor. That way you can see the drop in absolute
> performance etc..

Ok, here are some numbers. I should be able to get the rest during the 
week.
The averages are pretty close, so I figured I would include some 
percentile
numbers. PELT and WALT numbers are with schedutil. On average it seems 
we're
introducing about 0.5% overhead with the current additional accounting.

Test: perf bench sched messaging -g 50 -l 5000
+-------------+-------------+----------+----------+-----------+
| 4.8.2-walt  | Performance | Ondemand | PELT     | WALT(10ms)|
+-------------+-------------+----------+----------+---------- +
|             |             |          |          |           |
|  90th       | 17.077      | 17.088   | 17.088   | 17.159    |
|             |             |          |          |           |
|  96th       | 17.117      | 17.421   | 17.198   | 17.343    |
|             |             |          |          |           |
|  Average    | 16.910      | 16.916   | 16.937   | 16.962    |
|             |             |          |          |           |
+-------------+-------------+----------+----------+-----------+
| 4.8.2-raw   | Performance | Ondemand | PELT     | WALT(10ms)|
+-------------+-------------+----------+----------+-----------+
|             |             |          |          |           |
|  90th       | 16.919      | 17.1694  | 16.986   | 0         |
|             |             |          |          |           |
|  96th       | 16.980      | 17.364   | 17.052   | 0         |
|             |             |          |          |           |
|  Average    | 16.841      | 16.902   | 16.860   | 0         |
+-------------+-------------+----------+----------+-----------+