[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <2677b01d-899d-4f64-b17c-85033386a4d3@arm.com>
Date: Tue, 5 Dec 2023 14:47:58 +0000
From: Hongyan Xia <hongyan.xia2@....com>
To: Vincent Guittot <vincent.guittot@...aro.org>
Cc: Ingo Molnar <mingo@...hat.com>,
Peter Zijlstra <peterz@...radead.org>,
Dietmar Eggemann <dietmar.eggemann@....com>,
Juri Lelli <juri.lelli@...hat.com>,
Qais Yousef <qyousef@...alina.io>,
Morten Rasmussen <morten.rasmussen@....com>,
Lukasz Luba <lukasz.luba@....com>,
Christian Loehle <christian.loehle@....com>,
linux-kernel@...r.kernel.org
Subject: Re: [RFC PATCH 2/6] sched/uclamp: Simulate PELT decay in
util_avg_uclamp
On 04/12/2023 16:07, Vincent Guittot wrote:
> On Wed, 4 Oct 2023 at 11:05, Hongyan Xia <Hongyan.Xia2@....com> wrote:
>>
>> From: Hongyan Xia <hongyan.xia2@....com>
>>
>> Because util_avg_uclamp is not directly managed by PELT, it lacks the
>> nice property of slowly decaying to a lower value, resulting in
>> performance degredation due to premature frequency drops.
>
> That a very good reason for not doing this
This is not much different from util_guest and the "additive uclamp"
proposal in LPC 2023. The major difference is that I introduce a new
signal right beside util_avg (which then needs to have PELT behavior)
and they introduce signals in a way like util_est.
Now that I think about it, maybe my way is indeed too drastic, and maybe
the util_est way is better.
>>
>> Add functions to decay root cfs utilization and tasks that are not on
>> the rq. This way, we get the benefits of PELT while still maintaining
>> uclamp. The rules are simple:
>
> Nack. This just highlights that you are mixing different things and
> then trying to make it work.
>
> Please keep PELT out of uclamp_min/max
Well, like in my previous comment I think PELT is already a mixed signal
anyway, and treating it like it's not mixed with uclamp has already
shown many problems which need corner case code like 0 spare capacity
and uclamp filtering to make things work properly, and the code to fix
uclamp is still growing.
I will see if I can rework this series in a util_est style, and will
probably converge with the other two proposals, but fundamentally the
idea is that we don't have a pure PELT signal anyway. Achieving a PELT
value of X doesn't mean much if we don't know under what uclamp values
was it achieved, and having a clamped(X) on the side can be a very good
hint.
Hongyan
Powered by blists - more mailing lists