Re: [PATCH 1/3] sched/fair: Add tg_load_contrib cfs_rq decay checking

From: Vincent Guittot <vincent.guittot@linaro.org>
To: Odin Ugedal <odin@uged.al>
Cc: Ingo Molnar <mingo@redhat.com>,
	Peter Zijlstra <peterz@infradead.org>,
	Juri Lelli <juri.lelli@redhat.com>,
	Dietmar Eggemann <dietmar.eggemann@arm.com>,
	Steven Rostedt <rostedt@goodmis.org>,
	Ben Segall <bsegall@google.com>, Mel Gorman <mgorman@suse.de>,
	Daniel Bristot de Oliveira <bristot@redhat.com>,
	"open list:CONTROL GROUP (CGROUP)" <cgroups@vger.kernel.org>,
	linux-kernel <linux-kernel@vger.kernel.org>
Subject: Re: [PATCH 1/3] sched/fair: Add tg_load_contrib cfs_rq decay checking
Date: Tue, 25 May 2021 16:30:56 +0200	[thread overview]
Message-ID: <CAKfTPtCaZOSEzRXVN9fTR2vTxGiANEARo6iDNMFiQV5=qAA4Tw@mail.gmail.com> (raw)
In-Reply-To: <CAFpoUr0f50hKUtWvpTy221xT+pUocY7LXCMCo3cPJupjgMtotg@mail.gmail.com>

On Tue, 25 May 2021 at 12:34, Odin Ugedal <odin@uged.al> wrote:
>
> Hi,
>
> tir. 25. mai 2021 kl. 11:58 skrev Vincent Guittot <vincent.guittot@linaro.org>:
> > Could you give more details about how cfs_rq->avg.load_avg = 4 but
> > cfs_rq->avg.load_sum = 0 ?

> >
> > cfs_rq->avg.load_sum is decayed and can become null when crossing
> > period which implies an update of cfs_rq->avg.load_avg.  This means
> > that your case is generated by something outside the pelt formula ...
> > like maybe the propagation of load in the tree. If this is the case,
> > we should find the error and fix it
>
> Ahh, yeah, that could probably be described better.
>
> It is (as far as I have found out) because the pelt divider is changed,
> and the output from "get_pelt_divider(&cfs_rq->avg)" is changed, resulting
> in a different value being removed than added.

ok so IIUC, it happens during the adding/removing/propagating
entities' load in the cfs_rq.

>
> Inside pelt itself, this cannot happen. When pelt changes the load_sum, it
> recalculates the load_avg based on load_sum, and not the delta, afaik.
>
> And as you say, the "issue" therefore (as I see it) outside of PELT. Due to
> how the pelt divider is changed, I assume it is hard to pinpoint where the issue
> is. I can try to find a clear path where where we can see what is added
> and what is removed from both cfs_rq->avg.load_sum and cfs_rq->avg.load_avg,
> to better be able to pinpoint what is happening.
>
> Previously I thought this was a result of precision loss due to division and
> multiplication during load add/remove inside fair.c, but I am not sure that
> is the issue, or is it?

I don't think the precision looss is the problem because
adding/removing load in fair.c could truncate load_sum but it stays
sync with load_avg. I will have a llo to see if i can see something
weird

>
> If my above line of thought makes sense, do you still view this as an error
> outside PELT, or do you see another possible/better solution?
>
> Will investigate further.

Thanks

>
> Thanks
> Odin