small addition: another test - running pgbench from postgres against
postgres 7.1beta4 on the same partition with all three filesystems
on the same machines results in much more logic results: on both
systems the results are best when xfs is used as filesystem -
so looks like the below problem is very specific to dbench
and smp ...
t
p.s.: btw. i reproducable and constantly see about 15% better results
for the pgbench produced tpc numbers then running the data-
base on an xfs filesystem compared to ext2 and reiserfs
which looks pretty good i think - this is for 7.1beta4
- 7.0.x is worse because of the required fsyncs in it
which results in xfs being not as good as the others
(at least ext2 which i compared it to)
Thomas Graichen <graichen@xxxxxxxxxxxxx> wrote:
> i have done some simple dbench test with ext2, reiserfs and xfs - all
> with the same kernel, same partition and run multiple times ... i did
> them on one machine (1 p233mmx, 64mb, ide) and the results looked
> like i expected them to be: all three filesystems are in about the
> same class with xfs the last - but for xfs the io between the dbench
> clients are much better balanced (all end close to eachother at the
> end of the run) - so - as steve once said - this might account for the
> a bit lower xfs results ... ok - then i ran them on another machine
> (2 pII333, 128mb, ide) and on this smp machine i now get only about
> 1/4 of the performance of ext2 and reiserfs with xfs ... it's in this
> setup absolutely reproducable (~1.5mb/sec vs. ~5.5mb/sec) ... i'll
> try to check this on another machine too if i find the time on monday
> - but maybe someone else may try it too - any idea?
> t
> p.s.: all this is with 2.4.1-XFS from about a week ago and without
> any kio stuff on ide with udma2 enabled
> --
> thomas.graichen@xxxxxxxxxxxxxx
> innominate AG
> the linux architects
> tel: +49-30-308806-13 fax: -77 http://www.innominate.com
--
thomas.graichen@xxxxxxxxxxxxxx
innominate AG
the linux architects
tel: +49-30-308806-13 fax: -77 http://www.innominate.com
|