Debian Installer + Software RAID5 + LVM2 + ext3 = write performance of ?
 

Have a Debian Etch install which is showing some poor performance when writing to the drives. According to top, my IOWaits get into the 70% to 100% range while trying to write to the drives files of more than a few megabytes. Writing a file with gigabytes of data will put the server into this high IOWaits state for tens of minutes after the writing of the file displays as completed at the command line.

Setup is four 500GB SATA drives with Software RAID5 to bind them together and LVM2 to create separate logical volumes with EXT3 as the top level filesystem. This is running on an Opteron system with 4GB of RAM using the motherboard's SATA controller to connect the drives.

Reading through the Software RAID howto, I found the stride=nn option for mkfs.ext3. Unfortunately, I do not find anything confirming if the Debian installer recognizes the option should be used when creating a ext3 filesystem ontop of a software RAID. With the write scheme RAID5 uses, I could see the stride option not being used as a significant contributing factor to the poor write performance I'm seeing.

During my google searches, I found a fellow on another site having a similarly described problem asking the same question in the last year. However, he never received an answer to the Debian installer functionality question.

Just an idea: perhaps one (or worse: more then 1) of your disks might be failing. Not likely if they're quite new though...

Ext3 isn't the best of filesystems for dealing with large files. So the performance-hit you see may not be so unexpected. Check this comparision between the various filesystems on Wiki: http://en.wikipedia.org/wiki/Comparison_of_file_systems Also, I've read a review from someone, but didn't bookmark it and I can't find it now. But that review put XFS as the best suitable filesystem for dealing with large files, closely followed by JFS and on some distance ReiserFS.

Drives were delivered over the summer. Hardware failuring is always a possiblity, but I would expect to get some alerts in the logs from SMART or mdadm if any of the drives are. None have been seen so far.

Another possiblity I have run across is the size of the 'stripe cache'. I've found a couple references promoting making it larger to improve write performance under RAID5. Etch's default appears to be 256.

Things found while researching. Figuring they may be of use to others.

http://wiki.novell.com/index.php/File_System_Primer
http://wiki.novell.com/index.php/Linux_Data_Management
http://www.ubuntugeek.com/how-to-inc...rformance.html
http://oss.sgi.com/projects/xfs/
http://wiki.centos.org/HowTos/Disk_Optimization