RE: [RFC 0/1] Alternate history mechanism for the TEO governor

From: Doug Smythies
Date: Mon May 25 2020 - 14:32:16 EST


On 2020.05.21 04:09 Pratik Sampat wrote:
> On 17/05/20 11:41 pm, Doug Smythies wrote:
> > On 2020.05.11 Pratik Rajesh Sampat wrote:
> >> First RFC posting:https://lkml.org/lkml/2020/2/22/27
> > Summary:
> >
> > On that thread I wrote:
> >
> > > I have done a couple of other tests with this patch set,
> > > but nothing to report yet, as the differences have been
> > > minor so far.
> >
> > I tried your tests, or as close as I could find, and still
> > do not notice much difference.
>
> That is quite unfortunate. At least it doesn't seem to regress.

Yes, while I have not been able to demonstrate improvement,
I have not found any regression.

>
> Nevertheless, as Rafael suggested aging is crucial, this patch doesn't age
> weights. I do have a version with aging but I had a lot of run to run variance
> so I had refrained from posting that.
> I'm tweaking around the logic for aging as well as distribution of weights,
> hopefully that may help.

O.K. I am putting this testing aside for now.
I like the set of tests, as they really show the differences between menu
and teo governors well.

> >>
> >> Sleeping Ebizzy
> >> ---------------
> >> Program to generate workloads resembling web server workloads.
> >> The benchmark is customized to allow for a sleep interval -i
> > I found a Phoronix ebizzy, but without the customization,
> > which I suspect is important to demonstrate your potential
> > improvement.
> >
> > Could you send me yours to try?
>
> Sure thing, sleeping ebizzy is hosted here:
> https://github.com/pratiksampat/sleeping-ebizzy
>
> >
> > ebizzy (records per second, more is better)
> >
> > teo wtteo menu
> > 132344 132228 99.91% 130926 98.93%

O.K. yours is way different than what I was using.
Anyway, results still are not very different
between teo and wtteo. Some tests are showing a little difference
between above/below statistics [1]

[1] http://www.smythies.com/~doug/linux/idle/wtteo/ebizzy-interval/2_below.png

By the way, and likely not relevant, your sleeping-ebizzy test
seems extremely sensitive to the interval and number of threads.
It is not clear to me what settings I should use to try to re-create
your results. [2] is an interesting graph of records per second verses
intervals verses threads.

[2] http://www.smythies.com/~doug/linux/idle/wtteo/doug08/sleeping-ebizzy-records-intervals-threads.png