Re: Redo block size...

Holger Baer <holger.baer_at_science-computing.de> wrote in message news:<bu8moo$7e7$1_at_news.BelWue.DE>...
> Noons wrote:
> > Holger Baer <holger.baer_at_science-computing.de> wrote in message news:<bu62bf$qpm$1_at_news.BelWue.DE>...
> >
> >
> >>>What filesystem would you recommend for the datafiles/redos on Linux RH? I've heard that
> >>>both, ext2 and ext3 are suitable.
> >>>
> >
> >
> >>according to this link, ext3 it is:
> >>
> >>http://www.quest-pipelines.com/newsletter-v2/linux2.htm
> >
> >
> > <careful>
> > According to that link, you can NOT derive that any of the file
> > systems detailed there are better for redo.
> >
> > The benchmarks do NOT discriminate between database datafiles
> > (using a large block size typically >4K) and redo logfiles
> > using MUCH SMALLER block sizes.
> >
> > The results on that link assume that all writes in Oracle are of
> > the same db block size. Nothing could be more wrong.
> >
> >
> >>And in the Oracle on Linux-Course, Oracle recommends using of ext3.
> >
> >
> > That's my gut feel too, but it's only that: a feel.
> > Based on it being similar to ext2 plus having journaling
> > (which prevents fsck problems at boot time). Debatable.
> >
> > The link above has done nothing to confirm or deny which f/s
> > is better for redo.
> > </careful>
> >
>
> Thanks for your corrections. When I first read that article, I was
> just overwhelmed how especially the raw files and jfs subperformed
> and that probably got me carried away ;-)
>
> However, I felt a small correction to Marcin's plain ext2 is faster
> was called for; as usual it all depends.
>
> As a side note: My personal feeling is that since ext3 is the default
> filesystem for RH AS 2.1 (again, according to the Oracle course notes),
> Oracle just recommends what is the default anyway, and the mentioned
> link at least shows that it doesn't necessarily kill performance...
>
> But to know what is best, the OP will have to test for him self, I'll
> admit.
>
> Cheers,
>
> Holger

Slightly OT but interesting: http://kerneltrap.org/node/view/715

Maybe we can come up with a similar benchmark test using redo.

jg

--
@home.com is bogus.
http://www.signonsandiego.com/news/uniontrib/wed/business/news_1b14insure.html