Re: Linux filesystems & oracle performance

"Tanel Poder" <change_to_my_first_name_at_integrid.info> wrote in message news:3f68b2a6$1_1_at_news.estpak.ee...

> I'd be interested in the article, because my understanding is that ext3
> (which is basically ext2+journalling) only helps in case of server crash,
> that during bootup you don't have to make a full file system check. I
> understand that this actually doen't make anything go faster or IO to take
> less. And reiserf is AFAIK most effective when dealing with huge number of
> files (such often are webservers), since it has a tree like inode structure.

ext3 uses a default block size for the f/s that is more appropriate for databases that are page (or block) based. Around 4K.

Both ext2 (2K) and ext3 (4K) MAY have the default block size changed, but that is not mentioned in ANY of the articles mentioned.

As such I have to conclude the metrics were taken with the default block size of each f/s. Which means ext3 will be doing roughly half the number of I/O requests of ext2 for the same type and volume of db operation. That certainly seems to be the case of the TPC numbers in the Quest paper.

If the measurements had been repeated with BOTH files systems configured for exactly the same block size, I'd say there would be a lot less difference between ext2 and ext3, with ext2 potentially being faster. This can however be countered by more efficient interface to OS buffering of ext3.

I've installed ext3 with 8K block sizes in my Lunix box for the db filesystems and ext2 for all other file systems (os, programs, etc) with very good results. The box is dual booted to W2K and Lunix is streets ahead in I/O.

Reiserfs is good for file servers of multiple small files, but not really appropriate for databases where files are usually very large. Having said that, I've never gone into it in detail to find out if it is tunable.

The really surprising thing in the Quest paper is RAW. I guess the LVM used is a little bit more than just a "logical" volume manager. Or else an I/O request for a raw device by Oracle gets somehow intercepted by Lunix and "translated" into something else.

I've gotta download that kernel source for Lunix, this sounds quite interesting to dig into...

-- 
Cheers
Nuno Souto
wizofoz2k_at_yahoo.com.au.nospam