On Friday 2011-03-11 02:10, Dave Chinner wrote:
>On Thu, Mar 10, 2011 at 03:14:34PM +0100, Jan Engelhardt wrote:
>>
>> Was there something I missed?
>>
>> # xfs_info /
>> meta-data=/dev/md3 isize=256 agcount=32,
>> agsize=11429117 blks
>> = sectsz=512 attr=2
>> data = bsize=4096 blocks=365731739,
>> imaxpct=5
>> = sunit=0 swidth=0 blks
>> naming =version 2 bsize=4096 ascii-ci=0
>> log =internal bsize=4096 blocks=32768, version=2
>> = sectsz=512 sunit=0 blks, lazy-count=0
> ^^^^^^^^^^^^
>> realtime =none extsz=4096 blocks=0, rtextents=0
>
>You're using an old mkfs?
As mentioned, this is a preexisting fs. This was created in August 2009
using xfsprogs 2.10.1.
>At minimum, this should have lazy-count=1.
>I'm also wondering about the fact this is a MD device but there is
>no sunit/swidth set,
>Further - what is your storage configuration (e.g. what type of MD
>raid are you using) and is the filesystem correctly aligned to the
>storage? If you get these wrong, then nothing else you do will
>improve performance.
mdraid1 over two dumb SATA disks.
>and the agcount of 32 is not a default value,
Right. xfsprogs had just switched from agcount=16 to agcount=4
for its default value, which at that time seemed a little
uncomforting, given disks grow in size. So as much as I can recall,
I only set the agcount manually to 32 (1.5T/32=46G) for
when I once (like, 2007 or so) created an xfs it used 16
(250G/16=15G).
>What are your mount options - perhaps you've missed the fact that
>the new functionality requires the "delaylog" mount option to be
>added.
Per /proc/mounts:
/dev/md3 / xfs rw,relatime,attr2,nobarrier,noquota 0 0
>Mind you, that is not a magic bullet - if the operation is
>single threaded and CPU bound, delaylog makes no difference to
>performance, and with lazy-count=0 then the superblock will still be
>a major contention point and probably nullify any improvement
>delaylog could provide..
The question is.. is the writeout single-threaded?
Judging from there being only one xfs thread per block device,
that may seem to hold.
|