Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Date:   Mon, 17 Jan 2022 20:44:19 +0300
From:   Alexander Fomichev <fomichev.ru@gmail.com>
To:     Mel Gorman <mgorman@suse.de>
Cc:     linux-kernel@vger.kernel.org, dmaengine@vger.kernel.org,
        linux@yadro.com, Peter Zijlstra <peterz@infradead.org>
Subject: Re: [RFC] Scheduler: DMA Engine regression because of sched/fair
 changes
Message-ID: <20220117174419.hqj2mz2ctrjdq57d@yadro.com>
References: <20220112152609.gg2boujeh5vv5cns@yadro.com>
 <20220112170512.GO3301@suse.de>
 <20220117081905.a4pwglxqj7dqpyql@yadro.com>
 <20220117102701.GQ3301@suse.de>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20220117102701.GQ3301@suse.de>
Precedence: bulk

On Mon, Jan 17, 2022 at 10:27:01AM +0000, Mel Gorman wrote:
> > 1) You're right. When options "noverify=1" and "polling=1" are used.
> > then no performance reducing occurs.
> 
> How about just noverify=1 on its own? It's a stronger indicator that
> cache hotness is a factor.
> 

With "noverify=1 polled=0" the performance reduction is only 10-20%,
but still exists.

-----< v5.15.8-vanilla >-----
[17057.866760] dmatest: Added 1 threads using dma0chan0
[17060.133880] dmatest: Started 1 threads using dma0chan0
[17060.154343] dmatest: dma0chan0-copy0: summary 1000 tests, 0 failures 49338.85 iops 3157686 KB/s (0)
[17063.737887] dmatest: Added 1 threads using dma0chan0
[17065.113838] dmatest: Started 1 threads using dma0chan0
[17065.137659] dmatest: dma0chan0-copy0: summary 1000 tests, 0 failures 42183.41 iops 2699738 KB/s (0)
[17100.339989] dmatest: Added 1 threads using dma0chan0
[17102.190764] dmatest: Started 1 threads using dma0chan0
[17102.214285] dmatest: dma0chan0-copy0: summary 1000 tests, 0 failures 42844.89 iops 2742073 KB/s (0)
-----< end >-----

-----< 5.15.8-ioat-ptdma-dirty-fix+ >-----
[ 6183.356549] dmatest: Added 1 threads using dma0chan0
[ 6187.868237] dmatest: Started 1 threads using dma0chan0
[ 6187.887389] dmatest: dma0chan0-copy0: summary 1000 tests, 0 failures 52753.74 iops 3376239 KB/s (0)
[ 6201.913154] dmatest: Added 1 threads using dma0chan0
[ 6204.701340] dmatest: Started 1 threads using dma0chan0
[ 6204.720490] dmatest: dma0chan0-copy0: summary 1000 tests, 0 failures 52614.96 iops 3367357 KB/s (0)
[ 6285.114603] dmatest: Added 1 threads using dma0chan0
[ 6287.031875] dmatest: Started 1 threads using dma0chan0
[ 6287.050278] dmatest: dma0chan0-copy0: summary 1000 tests, 0 failures 54939.01 iops 3516097 KB/s (0)
-----< end >-----

> > 2) DMA Engine on certain devices, e.g. Switchtec DMA and AMD PTDMA, is
> > used particularly for off-CPU data transfer via device's NTB to a remote
> > host. In NTRDMA project, which I'm involved to, DMA Engine sends data to
> > remote ring buffer and on data arrival CPU processes local ring buffers.
> > 
> 
> Is there any impact of the patch in this case? Given that it's a remote
> host, the data is likely cache cold anyway.
> 

It's complicated. Currently we have a bunch of problems with the
project. So we do decomposition and try to solve them separately. Here
we faced the DMA Engine issue.

> > 4) Do you mean that with noverify=N and dirty patch, data verification
> > is performed on cached data and thus measured performance is fake?
> > 
> 
> I think it's the data verification going slower because the tasks are
> not aggressively migrating on interrupt. The flip side is other
> interrupts such as IO completion should not migrate the tasks given that
> the interrupt is not necessarily correlated with data hotness.
> 

It's quite strange, because dmatest substitutes verification time from
overall test time. I suspect measurement may be inaccurate.

> > 5) What DMA Engine enabled drivers (and dmatest) should use as design
> > pattern to conform migration/cache behavior? Does scheduler optimisation
> > conflict to DMA Engine performance in general?
> > 
> 
> I'm not familiar with DMA engine drivers but if they use wake_up
> interfaces then passing WF_SYNC or calling the wake_up_*_sync helpers
> may force the migration.
> 

Thanks for the advice. I'll try to check if this is a solution.


-- 
Regards,
  Alexander