Re: [QUESTION] Long read latencies on mixed rw buffered IO

From: Amir Goldstein <amir73il@gmail.com>
To: Dave Chinner <david@fromorbit.com>
Cc: Matthew Wilcox <willy@infradead.org>,
	"Darrick J. Wong" <darrick.wong@oracle.com>,
	linux-xfs <linux-xfs@vger.kernel.org>,
	Christoph Hellwig <hch@lst.de>,
	linux-fsdevel <linux-fsdevel@vger.kernel.org>
Subject: Re: [QUESTION] Long read latencies on mixed rw buffered IO
Date: Tue, 26 Mar 2019 05:44:34 +0200	[thread overview]
Message-ID: <CAOQ4uxh9gArhXZdSgOJgT12Ov9JYyPNoMf9H=h-Q-2vwcd8k=Q@mail.gmail.com> (raw)
In-Reply-To: <20190325234838.GC23020@dastard>

On Tue, Mar 26, 2019 at 1:48 AM Dave Chinner <david@fromorbit.com> wrote:
>
> On Mon, Mar 25, 2019 at 09:57:46PM +0200, Amir Goldstein wrote:
> > On Mon, Mar 25, 2019 at 9:40 PM Matthew Wilcox <willy@infradead.org> wrote:
> > >
> > > On Mon, Mar 25, 2019 at 09:18:51PM +0200, Amir Goldstein wrote:
> > > > On Mon, Mar 25, 2019 at 8:22 PM Matthew Wilcox <willy@infradead.org> wrote:
> > > > > On Mon, Mar 25, 2019 at 07:30:39PM +0200, Amir Goldstein wrote:
> > > > > > On Mon, Mar 25, 2019 at 6:41 PM Matthew Wilcox <willy@infradead.org> wrote:
> > > > > > > I think it is a bug that we only wake readers at the front of the queue;
> > > > > > > I think we would get better performance if we wake all readers.  ie here:
> > > >
> > > > So I have no access to the test machine of former tests right now,
> > > > but when running the same filebench randomrw workload
> > > > (8 writers, 8 readers) on VM with 2 CPUs and SSD drive, results
> > > > are not looking good for this patch:
> > > >
> > > > --- v5.1-rc1 / xfs ---
> > > > rand-write1          852404ops    14202ops/s 110.9mb/s      0.6ms/op
> > > > [0.01ms - 553.45ms]
> > > > rand-read1           26117ops      435ops/s   3.4mb/s     18.4ms/op
> > > > [0.04ms - 632.29ms]
> > > > 61.088: IO Summary: 878521 ops 14636.774 ops/s 435/14202 rd/wr
> > > > 114.3mb/s   1.1ms/op
> > > >
> >
> > --- v5.1-rc1 / xfs + patch v2 below ---
> > rand-write1          852487ops    14175ops/s 110.7mb/s      0.6ms/op
> > [0.01ms - 755.24ms]
> > rand-read1           23194ops      386ops/s   3.0mb/s     20.7ms/op
> > [0.03ms - 755.25ms]
> > 61.187: IO Summary: 875681 ops 14560.980 ops/s 386/14175 rd/wr
> > 113.8mb/s   1.1ms/op
> >
> > Not as bad as v1. Only a little bit worse than master...
> > The whole deal with the read/write balance and on SSD, I imagine
> > the balance really changes. That's why I was skeptical about
> > one-size-fits all read/write balance.
>
> You're not testing your SSD. You're testing writes into cache vs
> reads from disk. There is a massive latency difference in the two
> operations, so unless you use O_DSYNC for the writes, you are going
> to see this cache-vs-uncached performance unbalance. i.e. unless the
> rwsem is truly fair, there is always going to be more writer
> access to the lock because they spend less time holding it and so
> can put much more pressure on it.
>

Yeh, I know. SSD makes the balance better because of faster reads
from disk. Was pointing out that the worse case I am interested in is
on spindles. That said, O_DSYNC certainly does improve the balance
and gives shorter worse case latencies. However, it does not make the
problem go away. i_rwsem taken (even for 4K reads) takes its toll
on write latencies (compared to ext4).

Thanks,
Amir.