On Sat, 28 Aug 2010 19:50:24 +0400
Ilya Bakulin <webmas...@kibab.com> wrote:

> On Sat, 28 Aug 2010 13:09:12 +0400
> Ilya Bakulin <webmas...@kibab.com> wrote:
> 
> > Right now I'm running phybs -w /dev/ad7p1
> 
> Well, phybs finished.

I feel like a flooder today :-)

I've created gnop provider on top of the existing ad7p1:
kibab-desktop# gnop create -o 512 -S 4096 ad7p1

So, ad7p1.nop is shifted by 512 bytes and resides right on the beginning of the 
physical sector. And it has 4096 "sector" size.
Running phybs:

kibab-desktop# ./phybs -w /dev/ad7p1.nop 
   count    size  offset    step        msec     tps    kBps

   16384    8192       0   32768       11222     311   11679
   16384    8192    4096   32768       11836     295   11073

    8192   16384       0   65536       17432     223    7518
    8192   16384    4096   65536       18148     214    7222
    8192   16384    8192   65536       17811     218    7358

    4096   32768       0  131072       20800     196    6301
    4096   32768    4096  131072       12651     323   10360
    4096   32768    8192  131072       15040     272    8714
    4096   32768   16384  131072       15970     256    8206

    2048   65536       0  262144        8005     255   16372
    2048   65536    4096  262144        8152     251   16077
    2048   65536    8192  262144        7636     268   17163
    2048   65536   16384  262144        9372     218   13985
    2048   65536   32768  262144        9356     218   14008
For some reason, phybs begins with sector size 8192... I expected it to begin 
with 4096...

Let's force 4096:
kibab-desktop# ./phybs -l 4096 -w /dev/ad7p1.nop
   count    size  offset    step        msec     tps    kBps

   32768    4096       0   16384       15621     173    8390

   16384    8192       0   32768       11331     308   11566
   16384    8192    4096   32768       13045     268   10047

    8192   16384       0   65536       28344     137    4624
    8192   16384    4096   65536       22540     172    5815
    8192   16384    8192   65536       24530     158    5343

    4096   32768       0  131072       16221     252    8080
    4096   32768    4096  131072       13986     292    9371
    4096   32768    8192  131072       14736     277    8894
    4096   32768   16384  131072       14370     285    9121

Perfomance is excellent!

Notice, that for two subsequent phybs invocations there is big difference in 
timings for the same parameters.
For example, writing with block size=16384 at offset=0:
1st run: 7518 kBps,
2nd run: 4624 kBps.

This tells us that we cannot really compare these numbers with each other. The 
most important, however, is that they are significantly higher than misaligned 
writes. So using gnop is really a good solution.

-- 
Regards,
Ilya Bakulin
http://kibab.com
xmpp://kibab...@jabber.ru

Attachment: signature.asc
Description: PGP signature

Reply via email to