Re: [PATCH v3] mmc: rtsx: improve performance for multi block rw

[Ulf Hansson <ulf.hansson@linaro.org>]

On Thu, 10 Feb 2022 at 07:43, Ricky WU <ricky_wu@realtek.com> wrote:
>
>
>
> > -----Original Message-----
> > From: Ulf Hansson <ulf.hansson@linaro.org>
> > Sent: Monday, February 7, 2022 7:11 PM
> > To: Ricky WU <ricky_wu@realtek.com>
> > Cc: tommyhebb@gmail.com; linux-mmc@vger.kernel.org;
> > linux-kernel@vger.kernel.org
> > Subject: Re: [PATCH v3] mmc: rtsx: improve performance for multi block rw
> >
> > [...]
> >
> > > > > > >
> > > > > > > Do you have any suggestion for testing random I/O But we think
> > > > > > > random I/O will not change much
> > > > > >
> > > > > > I would probably look into using fio,
> > > > > > https://fio.readthedocs.io/en/latest/
> > > > > >
> > > > >
> > > > > Filled random I/O data
> > > > > Before the patch:
> > > > > CMD (Randread):
> > > > > sudo fio -filename=/dev/mmcblk0 -direct=1 -numjobs=1 -thread
> > > > > -group_reporting -ioengine=psync -iodepth=1 -size=1G -name=mytest
> > > > > -bs=1M -rw=randread
> > > >
> > > > Thanks for running the tests! Overall, I would not expect an impact
> > > > on the throughput when using a big blocksize like 1M. This is also
> > > > pretty clear from the result you have provided.
> > > >
> > > > However, especially for random writes and reads, we want to try with
> > > > smaller blocksizes. Like 8k or 16k, would you mind running another
> > > > round of tests to see how that works out?
> > > >
> > >
> > > Filled random I/O data(8k/16k)
> >
> > Hi Ricky,
> >
> > Apologize for the delay! Thanks for running the tests. Let me comment on
> > them below.
> >
> > >
> > > Before(randread)
> > > 8k:
> > > Cmd: sudo fio -filename=/dev/mmcblk0 -direct=1 -numjobs=1 -thread
> > > -group_reporting -ioengine=psync -iodepth=1 -size=1G -name=mytest
> > > -bs=8k -rw=randread
> > > mytest: (g=0): rw=randread, bs=(R) 8192B-8192B, (W) 8192B-8192B, (T)
> > > 8192B-8192B, ioengine=psync, iodepth=1
> > > result:
> > > Run status group 0 (all jobs):
> > >    READ: bw=16.5MiB/s (17.3MB/s), 16.5MiB/s-16.5MiB/s
> > > (17.3MB/s-17.3MB/s), io=1024MiB (1074MB), run=62019-62019msec Disk
> > stats (read/write):
> > >   mmcblk0: ios=130757/0, merge=0/0, ticks=57751/0, in_queue=57751,
> > > util=99.89%
> > >
> > > 16k:
> > > Cmd: sudo fio -filename=/dev/mmcblk0 -direct=1 -numjobs=1 -thread
> > > -group_reporting -ioengine=psync -iodepth=1 -size=1G -name=mytest
> > > -bs=16k -rw=randread
> > > mytest: (g=0): rw=randread, bs=(R) 16.0KiB-16.0KiB, (W)
> > > 16.0KiB-16.0KiB, (T) 16.0KiB-16.0KiB, ioengine=psync, iodepth=1
> > > result:
> > > Run status group 0 (all jobs):
> > >    READ: bw=23.3MiB/s (24.4MB/s), 23.3MiB/s-23.3MiB/s
> > > (24.4MB/s-24.4MB/s), io=1024MiB (1074MB), run=44034-44034msec Disk
> > stats (read/write):
> > >   mmcblk0: ios=65333/0, merge=0/0, ticks=39420/0, in_queue=39420,
> > > util=99.84%
> > >
> > > Before(randrwrite)
> > > 8k:
> > > Cmd: sudo fio -filename=/dev/mmcblk0 -direct=1 -numjobs=1 -thread
> > > -group_reporting -ioengine=psync -iodepth=1 -size=100M -name=mytest
> > > -bs=8k -rw=randwrite
> > > mytest: (g=0): rw=randwrite, bs=(R) 8192B-8192B, (W) 8192B-8192B, (T)
> > > 8192B-8192B, ioengine=psync, iodepth=1
> > > result:
> > > Run status group 0 (all jobs):
> > >   WRITE: bw=4060KiB/s (4158kB/s), 4060KiB/s-4060KiB/s
> > > (4158kB/s-4158kB/s), io=100MiB (105MB), run=25220-25220msec Disk stats
> > (read/write):
> > >   mmcblk0: ios=51/12759, merge=0/0, ticks=80/24154, in_queue=24234,
> > > util=99.90%
> > >
> > > 16k:
> > > Cmd: sudo fio -filename=/dev/mmcblk0 -direct=1 -numjobs=1 -thread
> > > -group_reporting -ioengine=psync -iodepth=1 -size=100M -name=mytest
> > > -bs=16k -rw=randwrite
> > > mytest: (g=0): rw=randwrite, bs=(R) 16.0KiB-16.0KiB, (W)
> > > 16.0KiB-16.0KiB, (T) 16.0KiB-16.0KiB, ioengine=psync, iodepth=1
> > > result:
> > > Run status group 0 (all jobs):
> > >   WRITE: bw=7201KiB/s (7373kB/s), 7201KiB/s-7201KiB/s
> > > (7373kB/s-7373kB/s), io=100MiB (105MB), run=14221-14221msec Disk stats
> > (read/write):
> > >   mmcblk0: ios=51/6367, merge=0/0, ticks=82/13647, in_queue=13728,
> > > util=99.81%
> > >
> > >
> > > After(randread)
> > > 8k:
> > > Cmd: sudo fio -filename=/dev/mmcblk0 -direct=1 -numjobs=1 -thread
> > > -group_reporting -ioengine=psync -iodepth=1 -size=1G -name=mytest
> > > -bs=8k -rw=randread
> > > mytest: (g=0): rw=randread, bs=(R) 8192B-8192B, (W) 8192B-8192B, (T)
> > > 8192B-8192B, ioengine=psync, iodepth=1
> > > result:
> > > Run status group 0 (all jobs):
> > >    READ: bw=12.4MiB/s (13.0MB/s), 12.4MiB/s-12.4MiB/s
> > > (13.0MB/s-13.0MB/s), io=1024MiB (1074MB), run=82397-82397msec Disk
> > stats (read/write):
> > >   mmcblk0: ios=130640/0, merge=0/0, ticks=74125/0, in_queue=74125,
> > > util=99.94%
> > >
> > > 16k:
> > > Cmd: sudo fio -filename=/dev/mmcblk0 -direct=1 -numjobs=1 -thread
> > > -group_reporting -ioengine=psync -iodepth=1 -size=1G -name=mytest
> > > -bs=16k -rw=randread
> > > mytest: (g=0): rw=randread, bs=(R) 16.0KiB-16.0KiB, (W)
> > > 16.0KiB-16.0KiB, (T) 16.0KiB-16.0KiB, ioengine=psync, iodepth=1
> > > result:
> > > Run status group 0 (all jobs):
> > >    READ: bw=20.0MiB/s (21.0MB/s), 20.0MiB/s-20.0MiB/s
> > > (21.0MB/s-21.0MB/s), io=1024MiB (1074MB), run=51076-51076msec Disk
> > stats (read/write):
> > >   mmcblk0: ios=65282/0, merge=0/0, ticks=46255/0, in_queue=46254,
> > > util=99.87%
> > >
> > > After(randwrite)
> > > 8k:
> > > Cmd: sudo fio -filename=/dev/mmcblk0 -direct=1 -numjobs=1 -thread
> > > -group_reporting -ioengine=psync -iodepth=1 -size=100M -name=mytest
> > > -bs=8k -rw=randwrite
> > > mytest: (g=0): rw=randwrite, bs=(R) 8192B-8192B, (W) 8192B-8192B, (T)
> > > 8192B-8192B, ioengine=psync, iodepth=1
> > > result:
> > > Run status group 0 (all jobs):
> > >   WRITE: bw=4215KiB/s (4317kB/s), 4215KiB/s-4215KiB/s
> > > (4317kB/s-4317kB/s), io=100MiB (105MB), run=24292-24292msec Disk stats
> > (read/write):
> > >   mmcblk0: ios=52/12717, merge=0/0, ticks=86/23182, in_queue=23267,
> > > util=99.92%
> > >
> > > 16k:
> > > Cmd: sudo fio -filename=/dev/mmcblk0 -direct=1 -numjobs=1 -thread
> > > -group_reporting -ioengine=psync -iodepth=1 -size=100M -name=mytest
> > > -bs=16k -rw=randwrite
> > > mytest: (g=0): rw=randwrite, bs=(R) 16.0KiB-16.0KiB, (W)
> > > 16.0KiB-16.0KiB, (T) 16.0KiB-16.0KiB, ioengine=psync, iodepth=1
> > > result:
> > > Run status group 0 (all jobs):
> > >   WRITE: bw=6499KiB/s (6655kB/s), 6499KiB/s-6499KiB/s
> > > (6655kB/s-6655kB/s), io=100MiB (105MB), run=15756-15756msec Disk stats
> > (read/write):
> > >   mmcblk0: ios=51/6347, merge=0/0, ticks=84/15120, in_queue=15204,
> > > util=99.80%
> >
> > It looks like the rand-read tests above are degrading with the new changes,
> > while rand-writes are both improving and degrading.
> >
> > To summarize my view from all the tests you have done at this point (thanks a
> > lot); it looks like the block I/O merging isn't really happening at common
> > blocklayer, at least to that extent that would benefit us. Clearly you have shown
> > that by the suggested change in the mmc host driver, by detecting whether the
> > "next" request is sequential to the previous one, which allows us to skip a
> > CMD12 and minimize some command overhead.
> >
> > However, according to the latest tests above, you have also proved that the
> > changes in the mmc host driver doesn't come without a cost.
> > In particular, small random-reads would degrade in performance from these
> > changes.
> >
> > That said, it looks to me that rather than trying to improve things for one
> > specific mmc host driver, it would be better to look at this from the generic
> > block layer point of view - and investigate why sequential reads/writes aren't
> > getting merged often enough for the MMC/SD case. If we can fix the problem
> > there, all mmc host drivers would benefit I assume.
> >
>
> So you are thinking about how to patch this in MMC/SD?
> I don't know if this method is compatible with other MMC Hosts? Or they need to patch other code on their host driver

I would not limit this to the core layer of MMC/SD. The point I was
trying to make was that it doesn't look like the generic block layer
is merging the sequential I/O requests in the most efficient way, at
least for the eMMC/SD devices. Why this is the case, I can't tell. It
looks like we need to do some more in-depth analysis to understand why
merging isn't efficient for us.

>
> > BTW, have you tried with different I/O schedulers? If you haven't tried BFQ, I
> > suggest you do as it's a good fit for MMC/SD.
> >
>
> I don’t know what is different I/O schedulers means?

What I/O scheduler did you use when running the test?

For MMC/SD the only one that makes sense to use is BFQ, however that
needs to be configured via sysfs after boot. There is no way,
currently, to make it the default, I think. You may look at
Documentation/block/bfq-iosched.rst, if you are more interested.

Kind regards
Uffe