2004.06_Filesystem Tuning-Getting the Most Performance Out of Your Filesystem.pdf

KNOW HOW

Filesystem Tuning

Hand-Formatted

Formatting programs not only take

care of installing filesystems. Like

well-organized toolkits, they also pro-

vide the right tool for the job. If you

need more power, you can tune the

mount options to squeeze the last

drop of performance out of Ext 3, JFS,

ReiserFS, or XFS. BY JÖRG REITTER

nowadays. JFS, ReiserFS, and XFS

handle large partitions really well.

The mkfs tool for JFS and XFS can format

a 1 Tbyte partition in no time at all. Reis-

erFS takes just 30 seconds. Ext 3 is a

special case, taking about an hour to do

the same job. That should be no surprise

as Ext 3 is really Ext 2 at its core.

Besides formatting tools there are any

number of utilities for each filesystem.

Some of them aim to improve perfor-

mance, others help troubleshoot errors.

This article looks at some of the most

important command line tools. The

boxes list the mount options, which can

be entered on the fly in the command

line as parameters of the mount com-

mand, or stored permanently in your

/etc/fstab file.

the root user, first launch tune2fs, from

the e2fsprogs [1] package, to update the

Ext 2 system:

Ext2/3 – Lots of Options

It is a good idea to add a journal to Ext 2,

thus updating to the Ext 3 filesystem. The

update is quite straightforward. But make

sure you have a good backup copy, just in

case.

To use Ext 3, you need to enable ker-

nel support for Ext 3. To use Ext 3 for

your root partition, you will also need to

compile the Ext 3 driver into your kernel.

If you boot Linux with an initial RAM

disk (initrd), you can implement the dri-

ver as a module. After ensuring that your

system fulfills the requirements, you can

start the upgrade process. Working as

tune2fs -j /dev/hda2

The -j flag simply creates the journal,

and stores the .journal file in the parti-

tion root. To complete the update,

change the filesystem type in /etc/fstab

from ext2 to ext3 and reboot. If you need

to mount as Ext 2, unmount the filesys-

tem, and specify the ext2 filesystem type

when remounting. If you need to make

the change permanent, revert to the orig-

inal filesystem type in /etc/fstab .

The bigger the filesystem, the more

time it takes to perform a filesystem

Table 1: XFS Sysctls

Option

Min value

Default

Max value

Explanation

fs.xfs.stats_clear

When set to 1 , deletes the statistics in /proc/fs/xfs/stat .

fs.xfs.sync_interval

30*HZ

60*HZ

Sets the interval at which xfssyncd writes the metadata to disk.

fs.xfs.error_level

Sets the error verbosity level. Generates more or less detailed reports and back-traces for filesystem

shutdowns. The following values for Errorlevel are defined: OFF: 0; LOW: 1, HIGH: 5.

fs.xfs.panic_mask

127

Debugging option. The BUG() is called for specific fields.

fs.xfs.irix_symlink_mode

Checks whether to create symbolic links with the default mode (0777), or if the mode is specified by the

umask preset (Irix mode).

fs.xfs.irix_sgid_inherit

Checks the files that users store in SGID directories. The option deletes the ISGID bit if the group ID for the

new file does not match the actual group ID or one of the additional group IDs of the parent directory.

fs.xfs.restrict_chown

Checks whether non-privileged users are allowed to use the chown command, to assign other users file

permissions.

vm.pagebuf.stats_clear

When set to 1 ,immediately deletes the statistics in /proc/fs/pagebuf/stat .

vm.pagebuf.flush_age

1*HZ

15*HZ

300*HZ

The age from which the content of the metadata buffer will be written out to disk.

vm.pagebuf.flush_int

HZ/2

30*HZ

The interval at which the list of meta data buffers is scanned.

June 2004

www.linux-magazine.com

Using filesystem tools

F ilesystem tests do not take too long

Filesystem Tuning

KNOW HOW

check, when the default interval has

expired. If a check takes too long to com-

plete, despite the journal, or if the

default interval is too short, you can

modify two options. -i allows you to set

the interval in days, weeks, or months, at

which e2fsck will analyze the filesystem.

You can also perform a check after a cer-

tain number of mounts. Set the -c option

to do so. Note that a value of 0 for both

options will disable filesystem checking.

ting. The predicted size can be bet-

ween 4 KBytes ( news ) and 4 MBytes

( largefile4 ). The following command

assumes an average value of 1 MByte:

using the journal. Then enter the follow-

ing command to create the journal for

hdb1 on the external device hda9 :

mkfs.ext3 -O /dev/hdb1 /dev/hda9

mkfs.ext3 -T largefile

Checking for Bad Blocks

Before you install the filesystem, you

should always check your hard disk for

errors. -c tells the mkfs.ext3 program to

check the disk for bad blocks. This is

always a good idea if your hardware has

done a fair amount of mileage. If you

want to know exactly how healthy your

disk is, specify the parameter twice. This

tells mkfs.ext3 to perform a read-write

test, rather than just a read-only text.

Unfortunately, this test does take twice

as long.

Mkfs.ext3 also accepts journal options.

Users can select to store the journal

internally in the superblock ( size=

1024 ), or externally on another device

( device=/dev/hdc2 ). If you opt for the

external variant, the journal must exist

before you use mkfs.ext3 to set up the

filesystem.

Tw o steps are required. First, you need

to format the partition where you will be

storing the journal. Ensure that this par-

tition uses the same block size (1, 2, or 4

KBytes) as the filesystem that will be

The filesystem on hdb1 uses the -J flag to

find the external journal:

tune2fs -J device=/dev/hda9

Formatting Permitted

Choose a small partition for test pur-

poses. If you do not specify any options,

the mkfs.ext3 formatting tool will supply

the missing parameters itself. That

makes it really easy to format a partition:

After specifying the options you need,

the next step is to simulate creating the

XFS: Mount Options

More detailed information on the XFS-

filesystem is available from /usr/src/linux/

Documentation/filesystems/xfs.txt

biosize= value : Sets the preferred size for

buffered I/O (default: 64 KByte), where value

represents the log of the required size.

ikeep :Keep empty Inode clusters instead of

assigning them to free disk pool.

logbufs= value : Sets the number of log

buffers in main memory. Values 2 through 8 .

Can increase performance while at the

same time impacting free memory.

logbsize= value : Sets the log buffer size.

Values 16, 32, 64, 128, 256 KBytes. Systems

with more than 32 MBytes RAM use 32

KByte by default.

logdev= device :Location of external log

( journal).

noalign :This option will not align data

allocations at the borders of stripe units.

noatime : Do not update the timestamp for

read access.

norecovery : Mounts the file system without

log recovery. The filesystem must be

mounted read-only, otherwise the mount

operation will fail.

osyncisosync : Updates the timestamp for

files where the O_SYNC flag is set. Perfor-

mance may be better without this flag, but

files will lose timestamp information in case

of a crash..

quota, usrquota, uqnoenforce : Enables user

quotas and optionally enforces the limits.

grpquota, gqnoenforce : Enables group

quotas and optionally enforces limits.

sunit= value and swidth= value : Sets the size

of the stripe unit in steps of 512 bytes (block

units). swidth must be set to a multiple of

sunit .

For Raid devices note that root can over-

write the information in the superblock, if

the disk layout is modified after creating the

filesystem. nouuid: No UUID check for multi-

ple mounting of filesystems. Useful for

mounting LVM snapshots.

mkfs.ext3 /dev/sdb2 E

The program will assume a block size

based on the partition size. To prevent

this, you need to set -b to specify the

block size (1, 2, or 4 KBytes) yourself:

mkfs.ext3 -b 2048 /dev/sdb2

mkfs.ext3 has a number of options to

deal with special requirements. -T allows

you to specify the typical file size on the

filesystem. The tool will set additional

optional parameters based on this set-

Ext3:Mount Options

More detailed information on the Ext 3

filesystem is available from /usr/src/linux/

Documentation/filesystems/ext3.txt .

data=journal :Stores the metadata in the

journal first, before writing data to the

filesystem (metadata and data journaling).

data=ordered :Stores the data on the file-

system first, before modifying the metadata.

data=writeback :Varies: sync() writes data

at set intervals. The metadata journal can

be modified before or following this

event.

journal=update : Update of the Ext 3 journals

to the current format.

journal= inum :Root can use inum to specify

the Inode number that represents the jour-

nal file. This is ignored if a journal already

exists.

noload : Does not load the journal.

minixdf :The df command acts like on a

Minix filesystem.

bsddf :The df command acts like on a

Berkeley Unix (BSD).

check=none, nocheck : Does not check

bitmaps on mounting.

debug :The kernel sends debugging informa-

tion to the syslog service.

errors=continue :Will continue to mount

despite filesystem errors.

errors=remount-ro :Will mount read-only in

case of filesystem error.

errors=panic :Will halt the machine on

filesystem error.

grpid, bsdgroups : Assigns the same group ID

as the parent object to an object.

nogrpid, sysvgroups :New objects are

assigned the group ID of the creator.

resuid= UID :The user ID of the user per-

mitted to use reserved blocks.

resgid= GID :The group ID of the group

permitted to use reserved blocks.

sb= Offset : Use an alternative superblock at

this location.

Note: Ext 2 ignores the quota options, such

as grpquota , noquota , quota , and usrquota ,

without issuing an error message.

www.linux-magazine.com

June 2004

KNOW HOW

Filesystem Tuning

filesystem by setting the -n flag. This is a

good way of finding out if your options

work. There are more options. For exam-

ple, the root user can reserve more than

five percent of the partition space for his

own use ( -m ), or disable some default

Ext 3 functions ( -O ).

Dumpe2fs outputs the content of the

superblock and the block groups. e2fsck,

performs filesystem checks. This does

not take long to complete on an Ext 3

system, thanks to the journal. The excep-

tion to this rule is the case where the

superblock is damaged, and needs to be

checked. With LVM, you can use

e2fsadm to reduce the size of the device;

the tool can handle online and offline

resizing. The Ext 2 dump and restore

utilities will work with Ext 3.

Some mount options are also aimed at

error handling, and are much discussed

topics in various mailing lists.

age contains various tuning and error

recovery tools. However, there are no

dump or restore programs, although it is

possible to resize the filesystem.

Although ReiserFS is the standard

filesystem for Suse Linux 9.0, there is no

supporting documentation on the discs.

Suse includes the formatting tool,

mkfs.reiserfs , but don’t expect to dis-

cover manpages or help files in the

/usr/share/doc directory.

The kernel documentation is just as

taciturn. Your only options are to surf the

Web, especially the developer, Hans

Reiser’s website [3]. This is worth doing

anyway, as the page has patches for the

brave, and mailing list subscribers pro-

vide answers to various questions.

The mkreiserfs command creates a

filesystem. There is an option to set the

block size, but ReiserFS v3.6 only sup-

ports a block size of 4096 Bytes. This

restriction will be lifted with the release

of version 4, which will be some time

this year. ReiserFS selects a few addi-

tional options, such as the hash

algorithm ( r5 , rupasov , tea ), or whether

to install version 3.5 or 3.6, itself, if you

fail to specify them.

Lack of JFS Tools

The mkfs.jfs tool creates a filesystem.

You can set the -c flag to check the disk

for bad blocks before formatting. JFS

also supports exporting the journal to

another device ( -J ):

mkfs.jfs -J journal_dev U

external-journal

will create the journal on the specified

device. The device option is then used to

bind the journal to the filesystem:

mkfs.jfs -J device=external- U

journal device

JFS Low-Fat Diet

The JFS port is advanced enough for test

purposes, but the lack of options means

it is not suitable for production use on

Linux. On the other hand, JFS produces

excellent benchmark results. The kernel

code has a debug interface for advanced

analysis. In user-space, admins can try

the jfs_fscklog tool, a jfsutils [2] tool that

extracts the content of the fsck service

logfile and outputs it to the console.

JFS does not provide a lot of user

options. For example, there is no way of

setting the block size. This said, the

manpages are well maintained, and pro-

vide useful examples, in contrast to the

documentation for other filesystems.

The device referred to at the end of this

line, represents the partition where the

filesystem will reside. You can either opt

for journal_dev or external-journal in a

command. If you need to retain OS/2

compatibility, set the -O flag to enable

case-insensitive support.

There are another three JFS applica-

tions: jfs_tune displays the volume label

and UUID. jfs_debugfs to replace the

data at a specific offset with a hex string.

jfs_fscklog extracts and displays the con-

tents of the fsck service logfile.

ReiserFS v3.6 is a stable filesystem that

has enough performance to satisfy

power users. The Reiserfsprogs [3] pack-

Full Speed Ahead with Reiser

The journaling options are restricted to

the size ( -s ) of the journal, storing the

file on an external device ( -j ), and the

offset at which the journal starts ( -o ).

Experts will typically specify the maxi-

mum transaction size. It is fairly safe to

ReiserFS: Mount Options

More detailed information on the ReiserFS-

filesystem is available from http://www.

namesys.com/mount-options.html

conv : Allows a ReiserFS v3.6 to mount a v3.5

filesystem using v3.6 code for any objects to

be created. The v3.5 utilities will not work on

this filesystem after this operation.

dontpanic : Ignores input/output errors while

editing the journal. This option is only avail-

able in raw mode.

hash=rupasov :Very quick hash function for

enormous directories and lots of unusual

filenames. Dangerous, as hash collisions are

to be expected.

hash=tea :Provides a high degree of random-

ness and a low probability of hash collisions.

Impacts performance, but should be used if

the r5 hash issues a EHASHCOLLISION mes-

sage.

hash=detect : Only makes sense when

mounting an older filesystem for the first

time. Recognizes the hash function and

writes it to the superblock.

hash=r5 : Standard hash.

hashed_relocation :Attempts to tune the

block allocator. May improve performance.

See also, no_unhashed_relocation .

no_unhashed_relocation :Attempts to tune

the block allocator. May improve perfor-

mance. See also, hashed_relocation .

noborder : Disables the border allocator algo-

rithm .May improve performance.

nolog : Disables journaling. May improve per-

formance. Caution: This automatically

disables quick recovery after a crash.

notail : Disables the tail feature. If you have

trouble with a program (such as Lilo) in nor-

mal operations, you should avoid storing

extremely small files directly in the tree with

ReiserFS.

pgc=low, high : Sets the activity level of the

Passive Garbage Collector. Only works with

the reiserfs-raw driver, which is not included

with the standard kernel.

raw : Mounts the filesystem in raw mode.

Not included with the standard kernel.

replayonly :Restores the transactions in the

journal without mounting the filesystem.

This option is only used by fsck under normal

circumstances.

resize= value :Expand a partition on the fly,

especially for LVM devices (Logical Volume

Management). Needs the Resizer utility from

the Reiserfs-utils package. Download from

the Namesys FTP Site at: ftp://ftp.namesys.

com/pub/reiserfsprogs

June 2004

www.linux-magazine.com

KNOW HOW

Filesystem Tuning

experiment with these options as mkreis-

erfs automatically selects a safe size, if

the user supplies an invalid option. If

you want to set a UUID (Universally

Unique Identifier), you can use the -u

option to specify this parameter.

possible to access to the filesystem.

rebuild-sb restores the superblock, if the

partition really does contain a filesystem.

tages are kernel based quota support,

and tools to match in the Xfsprogs [4]

package. Also, the XFS package has its

own dump and restore tools.

Resizing a Filesystem

The resizing tool, resize_reiserfs , is still

beta. The utility can expand or reduce a

filesystem’s size. This assumes that the

filesystem is not mounted during the

operation. If you do not specify a size ( -

s ), the tool will expand the filesystem as

far as possible. The command syntax is

quite straightforward at first glance.

Fully Featured

The simplest syntax for formatting a par-

tition is as follows:

Repairing ReiserFS

The Reiserfsck check and repair tool is

launched on booting the system. If a sys-

tem crashes, the tool checks the file-

system consistency, using the default

check setting and restores the latest sta-

tus from the journal. If check discovers a

serious error, such as an inconsistency of

the filesystem, it prompts the user to do

something about it. Check displays the

option required to repair the system on

the console.

It recommends fix-fixable for more

minor inconsistencies, such as invalid

directory entries. If the filesystem is

more seriously damaged, the user can

try the rebuild-tree option. This tells Reis-

erfsck to rebuild the whole filesystem.

This operation can take a lot longer, and

lead to total loss of the filesystem.

You need to enter the following if you

see a message such as read_super_block:

can’t find a reiserfs file system ,:

mkfs.xfs /dev/sda1

The following command places the jour-

nal on another file system:

resize_reiserfs -s -2G /dev/hdb2

mkfs.xfs -l logdev=/dev/sdb2, U

size=10000b /dev/sda1

reduces the filesystem on hdb2 by 2

GBytes from its original size. To expand

the filesystem by 2 GBytes, simply type -

2G instead of +2G . This is not quite as

simple as it seems. Users need to run

fdisk and modify the partition size to

reflect the filesystem size. When expand-

ing the filesystem size, the partition size

needs to be modified beforehand. When

reducing, the partition size can be modi-

fied later. The Reiserfstune program

allows users to define the journal and

maximum transaction sizes and to move

the journal to another block device.

Debugreiserfs is a troubleshooting

tool. If you call the tool without specify-

ing any options, it outputs the

superblock for a given device. The -p

option is quite interesting in this context.

It tells Debugreiserfs to extract the

filesystem metadata. The ReiserFS team

uses the option to remove errors:

formats the first partition on the first

SCSI hard disk, and creates the journal in

the second partition on the second disk.

XFS also has a very flexible approach

to block sizes, but not on Linux, unfortu-

nately. Block sizes between 512 Bytes

and 64 KBytes are possible, but the page

size defines the upper limit on Linux.

x86 Linux systems have a page size of 4

KBytes, but at least you have the option

to choose a smaller size.

XFS groups its options in categories

such as filesystem, inodes, journal, and

real-time. Users can assign location or

size values to these groups, allowing

them to define the size of the stripe unit

for a RAID system, or specify the per-

centage of variable inodes. mkfs.xfs -N

displays an overview.

Yo u can expand an XFS filesystem.

Xfs_growfs with the -d flag set maxi-

mizes the data section. -D allows admins

to specify the desired size in blocks. The

-r and -R options do the same thing for

the real-time section.

Experts have a lot to gain from using

XFS. The Xfs_db tool allows you to read

or write blocks. The blockuse option

returns the file using a given block. The

wealth of xfs_db options underlines the

fact that Linux can hold sway with major

Unix systems, thanks to XFS.

reiserfsck -rebuild-sb

The message typically indicates a cor-

rupt superblock, which makes it im-

JFS: Mount Options

More detailed information on the JFS-

filesystem is available from /usr/src/linux/

Documentation/ filesystems/jfs.txt .

iocharset= Name : Character set the system

should use, to convert from Unicode to

ASCII.

resize= number :Expands the volume on the

fly by this number of blocks.

nointegrity :Prevents writes to the journal.

Useful if you need to restore a volume

backup.

integrity :The journal logs changes to the

metadata. After a restore users should set

this and the nointegrity option on remount-

ing.

errors=continue :Will continue to mount

despite filesystem errors.

errors=remount-ro :Will mount read-only in

case of filesystem error.

errors=panic :Will halt the machine on

filesystem error.

debugreiserfs -p /dev/hda2/ | U

gzip -c > hda2.gz

Ext 2/3 users thinking of migrating to

ReiserFS should be aware of one restric-

tion. There is no dump and restore tool.

The ReiserFS FAQ recommends tar for

incremental backups. Suse users can

perform a full backup with YaST 2.

XFS: For the Exacting

With the exception of Ext 3, XFS creates

an impression of being most complete. It

not only provides numerous options, the

documentation is also exemplary. XFS is

the only filesystem that allows you to set

sysctls in the proc filesystem (see Table

1). XFS spoils power users with options

for various buffer sizes. Its other advan-

■

INFO

[1] E2fsprogs: http://e2fsprogs.sourceforge.

net/ext2.html

[2] Jfsutils: http://www-124.ibm.com/jfs

[3] Reiserfsprogs: http://www.namesys.com/

download.html

[4] Xfsprogs: http://oss.sgi.com/projects/xfs

June 2004

www.linux-magazine.com

Plik z chomika:

Inne pliki z tego folderu:

Inne foldery tego chomika: