Re: tuning for large files in xfs

[Nathan Scott <nathans@sgi.com>]

On Mon, May 22, 2006 at 11:57:44AM -0700, fitzboy wrote:
> I've got a very large (2TB) proprietary database that is kept on an XFS
> partition under a debian 2.6.8 kernel. It seemed to work well, until I
> recently did some of my own tests and found that the performance should
> be better then it is...
> 
> basically, treat the database as just a bunch of random seeks. The XFS
> partition is sitting on top of a SCSI array (Dell PowerVault) which has
> 13+1 disks in a RAID5, stripe size=64k. I have done a number of tests
> that mimic my app's accesses and realized that I want the inode to be
> as large as possible (which in an intel box is only 2k), played with su
> and sw and got those to 64k and 13... and performance got better.
> 
> BUT... here is what I need to understand, the filesize has a drastic
> effect on performance. If I am doing random reads from a 20GB file
> (system only has 2GB ram, so caching is not a factor), I get
> performance about where I want it to be: about 5.7 - 6ms per block. But
> if that file is 2TB then the time almost doubles, to 11ms. Why is this?
> No other factors changed, only the filesize.
> 
> Another note, on this partition there is no other file then this one
> file.
> 
> I am assuming that somewhere along the way, the kernel now has to do an
> additional read from the disk for some metadata for xfs... perhaps the
> btree for the file doesn't fit in the kernel's memory? so it actually
> needs to do 2 seeks, one to find out where to go on disk then one to
> get the data. Is that the case? If so, how can I remedy this? How can I
> tell the kernel to keep all of the files xfs data in memory?

Hi Tim,

[please CC this sort of question to xfs@oss.sgi.com]

Can you send xfs_info output for the filesystem and the output
from xfs_bmap -vvp on this file?

The file size has zero effect on performance, but the number and
layout of the files extents is very relevent.  How was this file
created?

cheers.

-- 
Nathan
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/