How can we customize linux file systems like ext2/3?

Question

Actually IOzone is a filesystem benchmark tool. The benchmark generates and measures a variety of file operations like sequential reading/writing, random reading/writing. So I want to customize my own Linux file system ext3 either for random reading(Database intensive or OLTP) or sequential reading(audio/video app) so that there will be better performance for specific application. And how can we change file systems parameters like block size, disk cache, buffer size etc to achieve better performance.

Answer 1

To change these things you either have to do them on filesystem creation with mke2fs ( which mkfs.ext4 for example uses ) or some of the tuning options are available after creation with with tune2fs. This Forum thread on Gentoo Forums was about ext3 tuning. Here's what I think are the important parts from that ( it's all copied with some formatting ) *( Also please be advised that data=journal disables some certain peformance enhancements in ext4)*

Using Directory Indexing

This feature improves file access in large directories or directories containing many files by using hashed binary trees to store the directory information. It's perfectly safe to use, and it provides a fairly substantial improvement in most cases; so it's a good idea to enable it:

tune2fs -O dir_index /dev/hdXY

This will only take effect with directories created on that filesystem after tune2fs is run. In order to apply this to currently existing directories, we must run the e2fsck utility to optimize and reindex the directories on the filesystem:

e2fsck -D /dev/hdXY 

Enable Full Journaling

By default, ext3 partitions mount with the 'ordered' data mode. In this mode, all data is written to the main filesystem and its metadata is committed to the journal, whose blocks are logically grouped into transactions to decrease disk I/O. This tends to be a good default for most people. However, I've found a method that increases both reliability and performance (in some situations): journaling everything, including the file data itself (known as 'journal' data mode). Normally, one would think that journaling all data would decrease performance, because the data is written to disk twice: once to the journal then later committed to the main filesystem, but this does not seem to be the case. I've enabled it on all nine of my partitions and have only seen a minor performance loss in deleting large files. In fact, doing this can actually improve performance on a filesystem where much reading and writing is to be done simultaneously. See this article written by Daniel Robbins on IBM's website for more information

In fact, putting /usr/portage on its own ext3 partition with journal data mode seems to have decreased the time it takes to run emerge --sync significantly. I've also seen slight improvements in compile time.

There are two different ways to activate journal data mode. The first is by adding data=journal as a mount option in /etc/fstab. If you do it this way and want your root filesystem to also use it, you should also pass rootflags=data=journal as a kernel parameter in your bootloader's configuration. In the second method, you will use tune2fs to modify the default mount options in the filesystem's superblock:

tune2fs -O has_journal -o journal_data /dev/hdXY

Please note that the second method may not work for older kernels. Especially Linux 2.4.20 and below will likely disregard the default mount options on the superblock. If you're feeling adventurous you may also want to tweak the journal size. (I've left the journal size at the default.) A larger journal may give you better performance (at the cost of more disk space and longer recovery times). Please be sure to read the relevant section of the tune2fs manual before doing so:

tune2fs -J size=$SIZE /dev/hdXY

Checking The Filesystem Options Using tune2fs

Well, now that we've tweaked our filesystem, we want to make sure those tweaks are applied, right? Surprisingly, we can do this options iusing the tune2fs utility quite easily. To list all the contents of the filesystem's superblock, we can pass the "-l" (lowercase "L") option to tune2fs:

tune2fs -l /dev/hdXY

Unlike the other tune2fs calls, this can be run on a mounted filesystem without harm, since it doesn't access or attempt to change the filesystem at such a low level.

This will give you a lot of information about the filesystem, including the block/inode information, as well as the filesystem features and default mount options, which we are looking for. If all goes well, the relevant part of the output should include "dir_index" and "has_journal" flags in the Filesystem features listing, and should show a default mount option of "journal_data".