Re: Bad SSD performance with recent kernels

[Shaohua Li <shaohua.li@intel.com>]

On Mon, 2012-01-30 at 08:36 +0100, Herbert Poetzl wrote:
> On Mon, Jan 30, 2012 at 03:22:38PM +0800, Shaohua Li wrote:
> > On Mon, 2012-01-30 at 08:13 +0100, Herbert Poetzl wrote:
> >> On Mon, Jan 30, 2012 at 11:17:38AM +0800, Shaohua Li wrote:
> >>> 2012/1/30 Wu Fengguang <wfg@linux.intel.com>:
> >>>> On Sun, Jan 29, 2012 at 02:13:51PM +0100, Eric Dumazet wrote:
> >>>>> Le dimanche 29 janvier 2012 à 19:16 +0800, Wu Fengguang a écrit :
> 
> >>>>>> Note that as long as buffered read(2) is used, it makes almost no
> >>>>>> difference (well, at least for now) to do "dd bs=128k" or "dd bs=2MB":
> >>>>>> the 128kb readahead size will be used underneath to submit read IO.
> 
> >>>>> Hmm...
> 
> >>>>> # echo 3 >/proc/sys/vm/drop_caches ;dd if=/dev/sda of=/dev/null bs=128k count=32768
> >>>>> 32768+0 enregistrements lus
> >>>>> 32768+0 enregistrements écrits
> >>>>> 4294967296 octets (4,3 GB) copiés, 20,7718 s, 207 MB/s
> 
> 
> >>>>> # echo 3 >/proc/sys/vm/drop_caches ;dd if=/dev/sda of=/dev/null bs=2M count=2048
> >>>>> 2048+0 enregistrements lus
> >>>>> 2048+0 enregistrements écrits
> >>>>> 4294967296 octets (4,3 GB) copiés, 27,7824 s, 155 MB/s
> 
> >>>> Interesting. Here are my test results:
> 
> >>>> root@lkp-nex04 /home/wfg# echo 3 >/proc/sys/vm/drop_caches ;dd if=/dev/sda of=/dev/null bs=128k count=32768
> >>>> 32768+0 records in
> >>>> 32768+0 records out
> >>>> 4294967296 bytes (4.3 GB) copied, 19.0121 s, 226 MB/s
> >>>> root@lkp-nex04 /home/wfg# echo 3 >/proc/sys/vm/drop_caches ;dd if=/dev/sda of=/dev/null bs=2M count=2048
> >>>> 2048+0 records in
> >>>> 2048+0 records out
> >>>> 4294967296 bytes (4.3 GB) copied, 19.0214 s, 226 MB/s
> 
> >>>> Maybe the /dev/sda performance bug on your machine is sensitive to timing?
> >>> I got similar result:
> >>> 128k: 224M/s
> >>> 1M: 182M/s
> 
> >>> 1M block size is slow, I guess it's CPU related.
> 
> >>> And as for the big regression with newer kernel than 2.6.38,
> >>> please check if idle=poll helps. CPU idle dramatically impacts
> >>> disk performance and even latest cpuidle governor doesn't help
> >>> for some CPUs.
> 
> >> here are the tests with idle=poll and after switching to 128k
> >> (instead of 1M) blocksize (same amount of data transferred)
> 
> >> kernel    ------------ read /dev/sda -------------
> >>           --- noop ---  - deadline -  ---- cfs ---
> >>           [MB/s]  %CPU  [MB/s]  %CPU  [MB/s]  %CPU
> >> --------------------------------------------------
> >> 3.2.2      45.82   3.7   44.85   3.6   45.04   3.4
> >> 3.2.2i     45.59   2.3   51.78   2.6   46.03   2.2
> >> 3.2.2i128 250.24  20.9  252.68  21.3  250.00  21.6
> 
> >> kernel    -- write ---  ------------------read -----------------
> >>           --- noop ---  --- noop ---  - deadline -  ---- cfs ---
> >>           [MB/s]  %CPU  [MB/s]  %CPU  [MB/s]  %CPU  [MB/s]  %CPU
> >> ----------------------------------------------------------------
> >> 3.2.2     270.95  42.6  162.36   9.9  162.63   9.9  162.65  10.1
> >> 3.2.2i    269.10  41.4  170.82   6.6  171.20   6.6  170.91   6.7
> >> 3.2.2i128 270.38  67.7  162.35  10.2  163.01  10.3  162.34  10.7
> 
> > What's 3.2.2i and 3.2.2i128? 
> 
> 3.2.2 ...... kernel with default options (bs=1M)
> 3.2.2i ..... kernel with idle=poll (bs=1M)
> 3.2.2i128 .. kernel with idle=poll (bs=128k)
> 
> > does idle=poll help?
> 
> doesn't look like, at least to me ...
what's your /sys/block/sdx/queue/max_sectors_kb? if you make it smaller,
does the performance increase? In my system, a smaller max_sectors_kb
makes bs=2M and bs=128k have similar performance, which makes me think
it's CPU doesn't catch up quickly after a request finishes.