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

[Peter Zijlstra]

On Fri, Oct 28, 2016 at 12:55:52AM -0700, Vikram Mulukutla wrote:
> On 2016-10-28 00:29, Peter Zijlstra wrote:
> >On Fri, Oct 28, 2016 at 12:10:39AM -0700, Vikram Mulukutla wrote:
> >>
> >>We propose Window-Assisted Load Tracking (WALT) as an alternative or
> >>additional load tracking scheme in lieu of or along with PELT, one
> >>that in our estimation better tracks task demand and CPU utilization
> >>especially for use cases on mobile devices. WALT was conceived by
> >>Srivatsa Vaddagiri to provide better perf-per-watt numbers on
> >>asymmetric CPU (frequency and/or IPC) implementations, (specifically
> >>on Qualcomm Snapgdragon chipsets running Android) and its metrics
> >>have been used to guide task placement and p-state selection (load
> >>balancing in CFS still uses PELT statistics). WALT is now present in
> >>the android-common kernel as well.
> >
> >And how come we only learn of it after its already shipping? Isn't that
> >arse backwards?
> 
> Yes, but also we were not confident that it would be close to being
> acceptable
> upstream since it was intricately tied to our HMP scheduler. However now
> that
> more parties including the folks at ARM are interested, and given that EAS
> exists and schedutil was merged into mainline, we felt it the right time
> to try and introduce the concept. In general we are seriously trying to get
> more things upstream and converge.

That's good to hear, great!

That said, I would much prefer to modify/augment existing PELT stuff
than to add an entirely new set of accounting on the side of it. But
yes, I (and others) are aware of the pain points you mentioned with
using the PELT signal for schedutil.

We'll have to see what we can come up with.