Re: XFS IO multiplication problem on centos/rhel 6 using hp p420i raid controllers

[Dennis Kaarsemaker <dennis.kaarsemaker@booking.com>]

On Thu, 2013-03-07 at 14:57 +1100, Dave Chinner wrote:
> On Wed, Mar 06, 2013 at 02:53:12PM +0100, Dennis Kaarsemaker wrote:
> > On Fri, 2013-03-01 at 06:40 +1100, Dave Chinner wrote:
> > > On Thu, Feb 28, 2013 at 03:12:16PM +0100, Dennis Kaarsemaker wrote:
> > > > Hello XFS developers,
> > > > 
> > > > I have a problem as described in the subject. If I read the xfs website
> > > > correctly, this would be a place to ask for support with that problem.
> > > > Before I spam you all with details, please confirm if this is true or
> > > > direct me to a better place. Thanks!
> > > 
> > > CentOS/RHEL problems can be triaged up to a point here. i.e. we will
> > > make an effort to pinpoint the problem, but we give no guarantees
> > > and we definitely can't fix it. If you want a btter triage guarantee
> > > and to talk to someone who is able to fix the problem, you need to
> > > work through the problem with your RHEL support contact.
> > 
> > Hi Dave,
> > 
> > Thanks for responding. We have filed support tickets with HP and Red Hat
> > as well, I was trying to parallelize the search for an answer as the
> > problem is really getting in the way here. So much so that I've offered
> > a bottle of $favourite_drink on a serverfault question to the one who
> > solves it, that offer applies here too :)
> > 
> > > Either way:
> > > 
> > > http://xfs.org/index.php/XFS_FAQ#Q:_What_information_should_I_include_when_reporting_a_problem.3F
> > 
> > A summary of the problem is this:
> > 
> > [root@bc290bprdb-01 ~]# collectl
> > #<----CPU[HYPER]-----><----------Disks-----------><----------Network---------->
> > #cpu sys inter  ctxsw KBRead  Reads KBWrit Writes   KBIn  PktIn  KBOut  PktOut 
> >    1   0  1636   4219     16      1   2336    313    184    195     12     133 
> >    1   0  1654   2804     64      3   2919    432    391    352     20     208 
> > 
> > [root@bc291bprdb-01 ~]# collectl
> > #<----CPU[HYPER]-----><----------Disks-----------><----------Network---------->
> > #cpu sys inter  ctxsw KBRead  Reads KBWrit Writes   KBIn  PktIn  KBOut  PktOut 
> >    1   0  2220   3691    332     13  39992    331    112    122      6      92 
> >    0   0  1354   2708      0      0  39836    335    103    125      9      99 
> >    0   0  1563   3023    120      6  44036    369    399    317     13     188 
> > 
> > Notice the KBWrit difference. These are two identical hp gen 8 machines,
> > doing the same thing (replicating the same mysql schema). The one
> > writing ten times as many bytes in the same amount of transactions is
> > running centos 6 (and was running rhel 6).
> 
> So what is the problem? it is writing too much on the on the centos
> 6 machine? Either way, this doesn't sound like a filesystem problem
> - the size and amount of data writes is entirely determined by the
> application.

For performing the same amount of work (processing the same mysql
transactions, the same amount of IO transactions resulting from them),
the 'broken' case writes ten-ish times as many bytes.

> > /dev/mapper/sysvm-mysqlVol /mysql/bp xfs rw,relatime,attr2,delaylog,allocsize=1024k,logbsize=256k,sunit=512,swidth=1536,noquota 0 0
> 
> What is the reason for using allocsize, sunit/swidth? Are you using
> them on other machines?

xfs autodetects them from the hpsa driver. They seem to be correct for
the raid layout (256 strips, 3 drives per mirror pool) and I don't seem
to be able to override them.

> And if you remove the allocsize mount option, does the behaviour on
> centos6.3 change? What happens if you set allocsize=4k?

The allocsize parameter has no effect. It was put in place to correct a
monitoring issue: due to mysql's access patterns, using the default
large allocsize on rhel 6 makes our monitoring report the filesystem as
much fuller than it actually is.

> > xfs_info:
> > 
> > [root@bc291bprdb-01 ~]# xfs_info /mysql/bp/
> > meta-data=/dev/mapper/sysvm-mysqlVol isize=256    agcount=16, agsize=4915136 blks
> >          =                       sectsz=512   attr=2
> > data     =                       bsize=4096   blocks=78642176, imaxpct=25
> >          =                       sunit=64     swidth=192 blks
> > naming   =version 2              bsize=4096   ascii-ci=0
> > log      =internal               bsize=4096   blocks=38400, version=2
> >          =                       sectsz=512   sunit=64 blks, lazy-count=1
> > realtime =none                   extsz=4096   blocks=0, rtextents=0
> > 
> > 
> > And for reference, xfs_info on centos 5:
> > 
> > [root@bc290bprdb-01 ~]# xfs_info /mysql/bp/
> > meta-data=/dev/sysvm/mysqlVol    isize=256    agcount=22, agsize=4915200 blks
> >          =                       sectsz=512   attr=0
> > data     =                       bsize=4096   blocks=104857600, imaxpct=25
> >          =                       sunit=0      swidth=0 blks, unwritten=1
> > naming   =version 2              bsize=4096  
> > log      =internal               bsize=4096   blocks=32768, version=1
> >          =                       sectsz=512   sunit=0 blks, lazy-count=0
> > realtime =none                   extsz=4096   blocks=0, rtextents=0
> 
> The only difference is that the centos 6 filesystem is configured
> with sunit/swidth. That affects allocation alignment, but nothing
> else. It won't affect IO sizes.
> 
> > Linux 2.6.18-308.el5 (bc290bprdb-01.lhr4.prod.booking.com) 	03/06/2013
> > 
> > Device:         rrqm/s   wrqm/s   r/s   w/s    rMB/s    wMB/s avgrq-sz avgqu-sz   await  svctm  %util
> > cciss/c0d0        6.95    27.09  7.72 270.96     0.19     2.90    22.71     0.07    0.25   0.22   6.00
> > cciss/c0d0p1      0.00     0.00  0.00  0.00     0.00     0.00    47.62     0.00    1.69   1.61   0.00
> > cciss/c0d0p2      0.00     0.00  0.00  0.00     0.00     0.00    14.40     0.00    4.07   4.06   0.00
> > cciss/c0d0p3      6.94    27.09  7.72 270.96     0.19     2.90    22.71     0.07    0.25   0.22   6.00
> > dm-0              0.00     0.00  0.45 32.85     0.01     0.13     8.34     0.02    0.49   0.07   0.24
> > dm-1              0.00     0.00  6.97 264.13     0.15     2.77    22.10     0.07    0.24   0.22   5.93
> 
> So, 8k IOs on centos 5...
> 
> > Linux 2.6.32-279.1.1.el6.x86_64 (bc291bprdb-01.lhr4.prod.booking.com) 	03/06/2013 	_x86_64_	(32 CPU)
> > 
> > Device:         rrqm/s   wrqm/s     r/s     w/s    rMB/s    wMB/s avgrq-sz avgqu-sz   await  svctm  %util
> > sda               0.00     3.60    6.00  374.40     0.06    44.18   238.17     0.11    0.28   0.16   6.08
> > dm-0              0.00     0.00    0.00    4.40     0.00     0.02     8.00     0.00    0.27   0.18   0.08
> > dm-1              0.00     0.00    6.00  373.20     0.06    44.11   238.56     0.11    0.28   0.16   6.04
> 
> And 128k IOs on centos 6. Unless there's a massive difference in
> file layouts, nothing in the filesystem would cause such a
> dramatic change in IO size or thoughput.

I see, so now I need to find out what's causing the larger average
request size. Would you happen to know a list of common causes?

-- 
Dennis Kaarsemaker, Systems Architect
Booking.com
Herengracht 597, 1017 CE Amsterdam
Tel external +31 (0) 20 715 3409
Tel internal (7207) 3409

_______________________________________________
xfs mailing list
xfs@oss.sgi.com
http://oss.sgi.com/mailman/listinfo/xfs