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 10:27:01 +0000
From:   Mel Gorman <mgorman@suse.de>
To:     Alexander Fomichev <fomichev.ru@gmail.com>
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: <20220117102701.GQ3301@suse.de>
References: <20220112152609.gg2boujeh5vv5cns@yadro.com>
 <20220112170512.GO3301@suse.de>
 <20220117081905.a4pwglxqj7dqpyql@yadro.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=iso-8859-15
Content-Disposition: inline
In-Reply-To: <20220117081905.a4pwglxqj7dqpyql@yadro.com>
User-Agent: Mutt/1.10.1 (2018-07-13)
Precedence: bulk

On Mon, Jan 17, 2022 at 11:19:05AM +0300, Alexander Fomichev wrote:
> On Wed, Jan 12, 2022 at 05:05:12PM +0000, Mel Gorman wrote:
> > On Wed, Jan 12, 2022 at 06:26:09PM +0300, Alexander Fomichev wrote:
> > > CC: Mel Gorman <mgorman@suse.de>
> > > CC: linux@yadro.com
> > > 
> > > Hi all,
> > > 
> > > There's a huge regression found, which affects Intel Xeon's DMA Engine
> > > performance between v4.14 LTS and modern kernels. In certain
> > > circumstances the speed in dmatest is more than 6 times lower.
> > > 
> > > 	- Hardware -
> > > I did testing on 2 systems:
> > > 1) Intel(R) Xeon(R) Gold 6132 CPU @ 2.60GHz (Supermicro X11DAi-N)
> > > 2) Intel(R) Xeon(R) Bronze 3204 CPU @ 1.90GHz (YADRO Vegman S220)
> > > 
> > > 	- Measurement -
> > > The dmatest result speed decreases with almost any test settings.
> > > Although the most significant impact is revealed with 64K transfers. The
> > > following parameters were used:
> > > 
> > > modprobe dmatest iterations=1000 timeout=2000 test_buf_size=0x100000 transfer_size=0x10000 norandom=1
> > > echo "dma0chan0" > /sys/module/dmatest/parameters/channel
> > > echo 1 > /sys/module/dmatest/parameters/run
> > > 
> > > Every test csse was performed at least 3 times. All detailed results are
> > > below.
> > > 
> > > 	- Analysis -
> > > Bisecting revealed 2 different bad commits for those 2 systems, but both
> > > change the same function/condition in the same file.
> > > For the system (1) the bad commit is:
> > > [7332dec055f2457c386032f7e9b2991eb05c2a0a] sched/fair: Only immediately migrate tasks due to interrupts if prev and target CPUs share cache
> > > For the system (2) the bad commit is:
> > > [806486c377e33ab662de6d47902e9e2a32b79368] sched/fair: Do not migrate if the prev_cpu is idle
> > > 
> > > 	- Additional check -
> > > Attempting to revert the changes above, a dirty patch for the (current)
> > > kernel v5.16.0-rc5 was tested too:
> > > 
> > 
> > The consequences of the patch is allowing interrupts to migrate tasks away
> > from potentially cache hot data -- L1 misses if the two CPUs share LLC
> > or incurring remote memory access if migrating cross-node. The secondary
> > concern is that excessive migration from interrupts that round-robin CPUs
> > will mean that the CPU does not increase frequency. Minimally, the RFC
> > patch introduces regressions of their own. The comments cover the two
> > scenarios of interest
> > 
> > +        * If this_cpu is idle, it implies the wakeup is from interrupt
> > +        * context. Only allow the move if cache is shared. Otherwise an
> > +        * interrupt intensive workload could force all tasks onto one
> > +        * node depending on the IO topology or IRQ affinity settings.
> > 
> > (This one causes remote memory accesses and potentially overutilisation
> > of a subset of nodes)
> > 
> > +        * If the prev_cpu is idle and cache affine then avoid a migration.
> > +        * There is no guarantee that the cache hot data from an interrupt
> > +        * is more important than cache hot data on the prev_cpu and from
> > +        * a cpufreq perspective, it's better to have higher utilisation
> > +        * on one CPU.
> > 
> > (This one incurs L1/L2 misses due to a migration even though LLC may be
> > shared)
> > 
> > The tests don't say but what CPUs to the dmatest interrupts get
> > delivered to? dmatest appears to be an exception that the *only* hot
> > data of concern is also related to the interrupt as the DMA operation is
> > validated.
> > 
> > However, given that the point of a DMA engine is to transfer data without
> > the host CPU being involved and the interrupt is delivered on completion,
> > how realistic is it that the DMA data is immediately accessed on completion
> > by normal workloads that happen to use the DMA engine? What impact does
> > it have to tbe test is noverify or polling is used?
> 
> Thanks for the comment. Some additional notes regarding the issue.
> 
> 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.

> 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.

> 3) I checked dmatest with noverify=0 on PTDMA dirver: AMD EPYC 7313 16-Core
> Processor/ASRock ROMED8-2T. The regression occurs on this hardware too.
> 

I expect it would be the same reason, the data is cache cold for the
CPU.

> 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.

> 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.

> 6) I didn't suggest RFC patch to real world usage. It was just a test
> case to find out a low speed cause.
> 

I understand but I'm relunctant to go with the dirty patch on the
grounds it would reintroduce another class of regressions.

-- 
Mel Gorman
SUSE Labs