>
> Hi,
>
NFS problem deleted - we will have to do some investigating there.
For metadata intensive runs like bonnie file creation, try building a
filesystem with a bigger log than default:
mkfs -t xfs -f -l size=32768b /dev/xxx
and mount using
mount -o logbufs=4
This will help, but not massively.
You are running into the journalling features of XFS big time here, try
using a larger number of files for reiserfs, it does not sync its
journal out to disk more than once a minute unless it fills it, and since
these are lots of empty files they take very little room in reiserfs.
ext2 is just manipulating memory and not doing disk I/O much at all in
this case.
In ext2 if you reset the system in the middle of this you will sit in
fsck for quite a while, in reiserfs I suspect you will find none of
your 30000 files exists. Notice that ext2 and reiserfs are both cpu
bound, xfs is disk bound. You pays your money (or not in the case of linux)
and you takes your choice.
Having said that, there are things which could be done to speed up the
delete case in xfs, they have been written down for several years now,
we just never had the internal bandwidth to implement them. If anyone is
feeling really brave......
Steve
>
> I also ran bonnie++ (www.coker.com.au/bonnie++/) on the 2.4.2-XFS cvs
> XFS partition (without any NFS this time), and then ran it again on
> the same partition with ext2 and ReiserFS. XFS had the best
> bandwidth (up to 30% more than the others when kio was used), but its
> delete performance was pretty bad - worse even than ext2. Random
> delete was kinda amazingly slow.
>
> ------Sequential Create------ --------Random Create--------
> -Create-- --Read--- -Delete-- -Create-- --Read--- -Delete--
> files /sec %CP /sec %CP /sec %CP /sec %CP /sec %CP /sec %CP
> ext2 16 365 99 +++++ +++ 9735 98 366 99 +++++ +++ 3119 99
> xfs 16 1473 39 +++++ +++ 1360 31 1486 39 +++++ +++ 261 7
> reiser 16 8868 100 +++++ +++ 11914 99 8957 100 +++++ +++ 10379 100
>
> this is with the default 16k files/dir, and the +++'s means it was
> too fast to measure (a good thing! :-)
>
> hope this helps... let me know if I can give you more info. The full
> bonnie++ info is at http://www.cita.utoronto.ca/~rjh/fs.bench.txt, but
> as always YMMV. It's a datpaoint anyway.
>
> cheers,
> robin
> --
> Robin Humble http://www.cita.utoronto.ca/~rjh/
|