On Fri, Oct 31, 2003 at 09:44:25PM -0500, Andrew Klaassen wrote:
> On Sat, Nov 01, 2003 at 12:14:45AM +0100, Simon Matter wrote:
> >
> > Did you try your setup with an external log?
>
> *Wow*, that made huge difference. From < .5MB/sec to >
> 10MB/sec. Still not as good as Ext3, but getting better.
>
> I've noticed the variance in amount of data going out to disk
> with XFS was much higher than with Ext3, too. Using the
> external log on XFS, as you suggested, and a default Ext3 (these
> both on software RAID 5, of course):
>
> Ext3:
> Mean: 17541.5MB/s
> Std. Dev: 2375.89MB/s
>
> XFS:
> Mean: 12960.4MB/s
> Std. Dev: 6535.16MB/s
>
> That's computed with a 5-second timeslices. In English, it
> means that XFS write rates are jumping all over the place, from
> .5MB/s to 20MB/s. Any reason why that would be?
IIRC, this is probably caused by the caching that md does
for the raid5 personality, which is sensitive to changes in
the size of blocks which are coming down from the filesystem
(if the size changes this cache is flushed, IIRC).
There are two kinds of writes which are smaller than the
(default) XFS 4K blocksize - log writes and SB/AGI/AGF/AGFL
(allocation group header structures) writes. Making the log
external eliminates the first kind (as will using a version
2 log with suitable stripe size).
The allocation group header structures are sector sized, so
you will be able to eliminate the second kind by going to a
4K sector size (assuming default 4K filesystem blocksize) -
try the -ssize=4k option to mkfs.xfs. I'd be interested in
hearing back as to whether this helps any...
ext3 will always be passing down constant blocksized IOs, I
believe, so isn't susceptible to this issue.
cheers.
--
Nathan
|