[DRBD-user] Full write speed on local disks and network but slow write-speed on drbd

Tom Fernandes anyaddress at gmx.net
Tue Oct 23 15:58:12 CEST 2012

Note: "permalinks" may not be as permanent as we would like,
direct links of old sources may well be a few messages off.


Hi Arnold,

------- Original-Nachricht --------
> Datum: Tue, 23 Oct 2012 13:44:35 +0200
> From: Arnold Krille <arnold at arnoldarts.de>
> To: drbd-user at lists.linbit.com
> Subject: Re: [DRBD-user] Full write speed on local disks and network but
> slow write-speed on drbd

> Hi,
> 
> I can't help with your issue but I can give some comments on your testing:
> 
> On Tuesday 23 October 2012 12:12:26 Tom Fernandes wrote:
> > tom at hydra04 [1522]:~$ sudo dd if=/dev/zero of=/mnt/test-io/test.dd bs=1M
> > count=5000
> > 5000+0 records in
> > 5000+0 records out
> > 5242880000 bytes (5.2 GB) copied, 13.8065 s, 380 MB/s
> 
> That test is rubbish. When you want to test with dd (which only gives you
> the write-rate for sequential access) you have to use oflag=sync, otherwise
> you will see the caching of fs and disk-layer. You should test with a
> multiple of  4M block-size, otherwise half your writing 1MB will be reading
> the other 3MB of the block.

The dd-command is not actually a test. It only shows how test.dd (which is 
used for testing the bandwidth) is produced. The dd-command itself is not part 
of the test. Sorry for not having been clear here. 

warm regards,


Tom

> 
> > tom at hydra04 [1523]:~$ sudo nc 10.0.0.2 1234 < /mnt/test-io/test.dd
> > # Summary:
> > # Piped    4.88 GB in 00h00m22.37s:  223.45 MB/second
> > ----------------------------------------------------------------------
> 
> I use iperf for this, easier commandline.
> 
> > ----- bonnie on local partition on hydra04 - hydra05 is the same
> ---------
> > tom at hydra04 [1498]:~$ sudo mount /dev/vg0/test-io  /mnt/test-io/; cd
> > /mnt/test-io/; time sudo bonnie -f -u root; cd -;  sudo umount
> /mnt/test-io/
> > Using uid:0, gid:0.
> > Writing intelligently...done
> > Rewriting...done
> > Reading intelligently...done
> > start 'em...done...done...done...done...done...
> > Create files in sequential order...done.
> > Stat files in sequential order...done.
> > Delete files in sequential order...done.
> > Create files in random order...done.
> > Stat files in random order...done.
> > Delete files in random order...done.
> > Version  1.96       ------Sequential Output------ --Sequential Input- --
> > Random-
> > Concurrency   1     -Per Chr- --Block-- -Rewrite- -Per Chr- --Block-- --
> > Seeks--
> > Machine        Size K/sec %CP K/sec %CP K/sec %CP K/sec %CP K/sec %CP 
> /sec
> > %CP
> > hydra04      96704M           495967  68 159106  20           546991  28
> > 727.7 11
> > Latency                         203ms     940ms               183ms    
> > 186ms Version  1.96       ------Sequential Create------ --------Random
> > Create--------
> > hydra04             -Create-- --Read--- -Delete-- -Create-- --Read--- -
> > Delete--
> >               files  /sec %CP  /sec %CP  /sec %CP  /sec %CP  /sec %CP 
> /sec
> > %CP
> >                  16  5425   6 +++++ +++ 20766  23 +++++ +++ +++++ +++
> +++++
> > +++
> > Latency               478us     707us     685us     339us      18us    
> > 724us
> >
> 1.96,1.96,hydra04,1,1350881245,96704M,,,,495967,68,159106,20,,,546991,28,72
> > 7.7,11,16,,,,,5425,6, +++++,+++,20766,23,+++++,+++,+++++,+++,+++++,
> > +++,,203ms,940ms,,183ms,186ms,478us,707us,685us,339us,18us,724us
> 
> Another interesting test is dbench which simulates windows-sessions with 
> access from word, excel and ms access. And the results of both bonnie and 
> dbench are far more realistic then of simple dd, because doing long
> sequential 
> reads is a special case only known in audio/video reproduction. And even
> there 
> it might be that several streams are mixed and thus several different
> files at 
> different positions are accessed interleaved.
> For office, email, database, webserver its never reading big chunks 
> sequentially, its aleays about reading a bit here and a bit there and be
> low 
> latency (aka seek time) with this, not high throughput in raw bandwidth.
> 
> Have fun,
> 
> Arnold



More information about the drbd-user mailing list