Hdparm:
The following optimizations techniques can be applied to (E)IDE drives to utilize 32 bit I/O and DMA more effectively.
- Get info: hdparm /dev/hda
- Get more info: hdparm -I /dev/hda
- Test Speed: hdparm -tT /dev/hda
- Increase Speed: hdparm -d1 -c3 -u1 /dev/hda
- -d1: Turn on DMA (0: off)
- -c3: IDE chipset support
- 0: Disable
- 1: Enable
- 3: 32 bit transfer
- -u1: Get/set interrupt umask. Driver ignores other interrupts during processing of disk interrupts. (0: off)
- Test Speed Again: hdparm -tT /dev/hda
File: /etc/sysconfig/harddisks:
USE_DMA=1Turn off hdparm DMA access: hdparm -d0 /dev/hda
EIDE_32BIT=1
Installation: RH/CentOS: yum install hdparm
Links:
Mount "noatime":
If the drive is mounted with the directive "noatime", then the
access time for the file is not updated in the inode table when the
file is read. This increases performance because it reduces time to
seek and time of writing the access time to the drive.
This is done most commonly for read only file systems such as on many
news servers.
Example: mount -o noatime,ro,dev,nouser -t ext3 /dev/hdc /mnt/news1
Ramdisk:
Go even faster by using ramdisk - Memory mapped as a disk.
See: How to use a Ramdisk for Linux
Also see: /usr/src/linux-2.4/Documentation/ramdisk.txt
Tools:
- Measure hared drive performance with IBM Drive Fitness Test (DFT) utility. Create a boot floppy with DFT image (use dd) and boot. Measures SCSI and IDE performance as well as some repair utilities.
- IoMeter: Server I/O benchmark tool.
Direct kernel to specify processor affinity for a process:
Set (or retrieve) the CPU affinity. The scheduler attempts to keep the process on the specified CPU. This is especially effective for threaded applications to improve cache coherency on SMP computers. It is also useful for software licensed to a CPU or for a single CPU system when it is in fact running on an SMP system. (i.e. Oracle)- taskset: retrieve or set a processes's CPU affinity
(Part of package "schedutils") - sched_setaffinity / sched_getaffinity: set and get a process's CPU affinity mask
Kernel enhancements to optimize SCSI for scanner access:
- See man page for sane: man sane-scsi
- Excerpt from man page:
First, make sure your kernel has SCSI generic support enabled. In ``make xconfig'', this shows up under ``SCSI support->SCSI generic support''. To keep scanning times to a minimum, it is strongly recom- mended to use a large buffer size for the generic SCSI driver. By default, Linux uses a buffer of size 32KB. This works, but for many cheaper scanners this causes scanning to be slower by about a factor of four than when using a size of 127KB. Linux defines the size of this buffer by macro SG_BIG_BUFF in header file /usr/include/scsi/sg.h. Unless a system is seriously short on memory, it is recommended to increase this value to the maximum legal value of 128*1024-512=130560 bytes. After changing this value, it is necessary to recompile both the kernel (or the SCSI generic module) and the SCSI backends. AUTHOR David Mosberger 13 May 1998
Kernel enhancements to increase the per process limit of open files:
Change the definition of NR_OPEN in local files:
- /usr/src/linux/include/linux/limits.h (Local file: kernel 2.2 RH 7.0-) (default 1024 typ.)
- /usr/include/limits.h (Local file) (default 1024 typ.)
- /usr/src/linux/include/linux/fs.h (Local file: kernel 2.2 RH 7.0-)
The default Red Hat Kernel comes with as much stuff compiled into the Kernel as they could fit. This allows it to support a vast array of devices and system needs. If you must add support for something new or if you wish to streamline your kernel with only what is needed so that it uses less memory, operates faster, and is optimized for your processor, then one must recompile the kernel.
To find out about the latest kernel, issue the command: finger @finger.kernel.org
Typically the source code as installed by the package manager installation is located in:
- Red Hat / Fedora:
- /usr/src/kernels/2.6.9-55.0.12.EL-i586 - It is now typical that the kernel image comes with the kernel source (Red Hat Enterprise / CentOS 4) all in the "kernel" RPM.
- /usr/src/redhat/SOURCES/ - patches and kernel will be placed here by kernel source RPM. (Fedora Core 3)
(bzip2 -d linux-2.6.9.tar.bz2, tar xf linux-2.6.9.tar) - /usr/src/linux-2.4/ - (Soft linked to /usr/src/linux-2.4.xx) Kernel 2.4. (Red Hat 7.1)
- /usr/src/linux/ - (Soft linked to /usr/src/linux-2.2.xx) Kernel 2.2 and older. (Red Hat 6.x, 7.0)
- Ubuntu / Debian:
- /usr/src/linux - (Soft linked to /usr/src/linux-source-2.6.15/) Kernel 2.6 (Ubuntu 6.11)
(install package: apt-get install linux-source will install the package linux-source-2.6.15) - linux-headers-2.6.15-27/
(install package: apt-get install linux-headers will install the package linux-headers-2.6.15-27)
- /usr/src/linux - (Soft linked to /usr/src/linux-source-2.6.15/) Kernel 2.6 (Ubuntu 6.11)
If you get new kernel source download to a new directory, create a soft linkt to /usr/src/linux/. i.e. /usr/src/linux-2.X.XX then ln -s /usr/src/linux-2.X.XX /usr/src/linux
Typically the kernel RPM installation (kernel-source-XXX.i386.rpm) will install the source and generate the appropriate links.
Pre-requisite packages for building the kernel from source:
(see: /usr/src/linux/Documentation/Changes for the software versions required and how to check them)
- Linux kernel 2.6:
- gcc, GNU make, binutils, util-linux, module-init-tools, e2fsprogs, jfsutils, reiserfsprogs, xfsprogs, pcmcia-cs, quota-tools, PPP, isdn4k-utils, nfs-utils, procps, oprofile
- Linux kernel 2.4:
- gcc, GNU make, binutils, util-linux, modutils, e2fsprogs, reiserfsprogs, pcmcia-cs, PPP, isdn4k-utils
Ubuntu 6.11 / Kernel 2.6:
The source code for Ubuntu installations is obtained and built into a Debian package by:
- Install Ubuntu prerequisites: apt-get install build-dep build-essential fakeroot kernel-package linux-kernel-headers linux-headers-`uname -r` module-init-tools linux-image linux-386 linux-kernel-devel module-init-tools iptables libdrm2 autofs ncurses-dev
- apt-get install linux-source
This will also download the specific kernel source. In this case linux-source-2.6.15 - Source location: /usr/src/linux-source-2.6.15.tar.bz2
- cd /usr/src/
- tar xjvf /usr/src/linux-source-version-number-here.tar.bz2
- cd linux-source-version-number-here
- cp -vi /boot/config-`uname -r` .config
(config file for existing configuration) - Customize kernel config: make menuconfig or xconfig
- Create a Debian ".deb" kernel package: fakeroot make-kpkg --initrd --revision=custom.1.0 kernel_image
Where "mykernel.1.0" is any label you want to apply.
or make-kpkg --initrd --append-to-version=mykernel kernel_image kernel_headers - Install Debian package: dpkg -i kernel-image-XXXXX.deb
Perform the following steps: (as root)
-
Setup:
- The source may already be installed as part of the installation or
from an RPM (See /usr/src/).
OR
You may also download the latest source from http://www.kernel.org/pub/linux/kernel/.
Move the downloaded file linux-2.X.XX.tar.gz to /usr/src/.
Uncompress file: tar -xzf linux-2.X.XX.tar.gz which will install everything in a subdirectory linux/. (To view contents use command tar -tzf linux-...tar.gz)
If using the file with the bz2 extension: bunzip2 linux-2.X.XX.tar.bz2 - cd /usr/src/linux/
- less README (Read the documentation. It tells you everything. Use space bar to view next page)
(Local file link to: /usr/src/linux/README) - Update /usr/include/
NOTE: PERFORM THIS STEP FOR KERNEL 2.2.x AND OLDER. DO NOT DO THIS FOR KERNEL 2.4 or 2.6.
The header files for Kernel 2.4/2.6 and the C library header files may no longer match and thus must remain separated.- cd /usr/include
- rm -fr asm linux scsi
- ln -s /usr/src/linux/include/asm asm
- ln -s /usr/src/linux/include/linux linux
- ln -s /usr/src/linux/include/scsi scsi
Note: The sym link is optional and some don't do it due to potential differences in the header files used by glibc and the kernel.
- cd /usr/src/linux (or /usr/src/linux-2.4)
- make distclean (Cleans up junk from old compiles) (Warning: removes .config)
- make mrproper (build the source tree)
At this time the sources are correctly installed. This also resets the /usr/src/linux/.config file to the system default.
SuSE setup:
- For SuSE (9.3), there are two kernel-source RPMs. One RPM resides on the "Source" DVD-2 and the other resides on the "Installation" DVD-1. Use the kernel-source RPM found on the "Installation" DVD-1. This one will register in the RPM database and creates the directories expected
Check with the command: rpm -q kernel-source
Kernel source code location: /usr/src/linux
Docs location: /usr/share/doc/packages/kernel-source - The SuSE kernel-source RPM will not produce a ".config"
file for the kernel compilation.
Produce the default ".config" file by: zcat /proc/config.gz > /usr/src/linux/.config
- The source may already be installed as part of the installation or
from an RPM (See /usr/src/).
-
Configuration:
Use one of the following tools to create the .config file. This gives you the chance to choose what goes into the kernel. You can choose support for many of the latest capabilities, device drivers, and can tune the kernel for particular uses. i.e. you can also optimize the compilation for the Pentium. (Default for RH 5.2 was a 386 with no math co-processor!) The default config file used to compile the default distribution kernel can be found in the source directory. (i.e. Fedora Core 3: /usr/src/redhat/SOURCES/kernel-2.6.9-i686.config, /usr/src/redhat/SOURCES/kernel-2.6.9-x86_64-smp.config, ... and all the others for the various hardware platforms.)
Your kernel must support your hardware. During the kernel configuration process, you will have to know what hardware support is needed in the kernel. To view the system hardware use the commands:
- PCI bus attached hardware: /sbin/lspci
- CPU information: cat /proc/cpuinfo
Pick one of the following:
- make config (Bash shell script)
- make menuconfig (uses text window curses)
- make xconfig (uses tck and wish) - recommended due to online help feature and intuitive interface. Save configuration to file: .config
- make oldconfig Build a configuration file based on defaults found in current .config file.
menuconfig Notes: "< > options can be compiled as a module.
- <*> = Compile into kernel (y)
- <M> = Compile as a module (m)
Recommendation for Pentium owners: General Setup choose :
- Kernel math emulations: N
- Processor type: Pentium
- Development, experimental or kernel debugging features: N
- If using netfilter capabilities, do not enable the kernel fast-switching option. That code will permits fast switching at a lower level in the IP stack than netfilter and thus will bypass the netfilter capabilities.
- Note: When requiring gobs of memory (> 1Gb), there is a kernel compile option which sets the limit on the amount of user memory one can use: "Processor type and features / Maximum Virtual Memory".
-
Compile:
Next: (from /usr/src/linux)- make dep (make dependencies. Older kernels only. Attempt and if not needed, make will return a message stating that this step is un-necessary. Not required for 2.6+ kernels.)
- make clean (At this point you have all the source gathered)
- Choose one of the following to create the new kernel and then move it to /boot/vmlinux:
- make zImage (compressed image)
- make bzImage (very compressed image good for recovery floppies - slower to boot because of time to uncompress)
- make image
- make bzdisk (Boot floppy with no LILO)
-
Install:
- mv /usr/src/linux/arch/i386/boot/zImage /boot/vmlinuz-release
ln -s /boot/vmlinuz-release /boot/vmlinuz
(If "make zImage" was chosen)
or
mv /usr/src/linux/arch/i386/boot/bzImage /boot/vmlinuz-release
ln -s /boot/bzImage-release /boot/vmlinuz
(If "make bzImage" was chosen)
Also move/copy the file /usr/src/linux.../System.map to /boot/System.map-2.X.XX
OR
make install
(Preferred. Potential Pitfall: On my SCSI system I had to execute this command after issuing the mkinitrd command below.) - Modules documentation: (not that helpful)
- make modules
- rm /lib/modules/Kernel_release (i.e. /lib/modules/2.2.16-3 )
- ln -s /lib/modules/new_modules_directory /lib/modules/Kernel_release
- make modules_install
Generates the file: /lib/modules/2.X.XX/modules.dep
- mv /usr/src/linux/arch/i386/boot/zImage /boot/vmlinuz-release
-
Configure Boot Loader:
(Grub or Lilo)- Configure grub:
File: /etc/grub.confdefault=1 - Kernel to boot (2nd in list) timeout=5 splashimage=(hd1,0)/boot/grub/splash.xpm.gz - Boot image location (hd1, 1st partition) hiddenmenu title Fedora Core (2.6.12-1.1372_FC3) root (hd1,0) - Root of GRUB reference. Same as /boot kernel /boot/vmlinuz-2.6.12-1.1372_FC3 ro root=LABEL=/ rhgb quiet initrd /boot/initrd-2.6.12-1.1372_FC3.img title Fedora Core (2.6.11-1.14_FC3) root (hd1,0) kernel /boot/vmlinuz-2.6.11-1.14_FC3 ro root=LABEL=/ rhgb quiet initrd /boot/initrd-2.6.11-1.14_FC3.img
Description:- default: Default kernel to boot. All options displayed but one will boot by default.
- timeout: Time for user to select kernel. After "timeout" seconds, default kernel chosen and boot begins.
- splashimage: Background image shown by Grub and location (hard drive hd1 and "0" or first partition (/boot).
- hiddenmenu: I recommend commenting this out while compiling and developing kernels. This option takes way the user's ability to choose which kernel to boot. You always want the ability to choose a known reliable kernel just in case the one you just built is a disaster.
- root: Grub root directory. In this example, drive hd1 and "0" or first partition (/boot).
- kernel: Kernel to boot, read only. Note that this is the same as the result of cat /proc/cmdline
- initrd: Initial ramdisk for a Linux format boot image. Sets parameters in the Linux setup area in memory.
- Grub Home page
- grub shell man page.
- Configure lilo:
Lilo must point to the new kernel. Edit /etc/lilo.conf.
Add a new image statement to point to the new kernel.
Keep the old as backup in case you need to boot it.
SCSI systems must perform an additional step to load drivers upon system boot:
- Execute the following command which creates the file used to let the system boot:
mkinitrd /boot/initrd-2.2.16-3.img 2.2.16-3
The second argument is the name of the sub-directory of the modules under the directory /lib/modules/. (In this case /lib/modules/2.2.16-3/This step is required by systems using modular SCSI support. If the SCSI support is compiled into the kernel and is NOT a module, then this step may not be necessary. The file /boot/initrd-2.2.16-3.img is specified as the RAM disk image used before the file system is available. (initrd = initial ram disk)
Run /sbin/lilo -v to configure the master boot record with data from lilo.conf.
Sample lilo.conf:
boot=/dev/sda
map=/boot/map - Location of the kernel used for boot install=/boot/boot.b prompt timeout=50 linear - Specific to SCSI configurations default=linux image=/boot/vmlinuz-2.2.16-3 - New kernel label=linux - Name to be displayed by Lilo boot manager initrd=/boot/initrd-2.2.16-3.img - Specific to SCSI boot drives. read-only root=/dev/sdb6 image=/boot/vmlinuz-2.2.14-5.0 label=OldLinux initrd=/boot/initrd-2.2.14-5.0.img - SCSI specific. read-only root=/dev/sdb6When the system boots and gives you the "Lilo:" prompt, you can type "?" to see your boot choices as defined by the above labels in lilo.conf. Type the label name of the kernel or system you wish to boot. A carriage return will choose the default.
- Execute the following command which creates the file used to let the system boot:
For a full list of boot parameters see: bootparam man page
Also see the Boot Prompt Howto.
- Configure grub:
Kernel version verification: uname -rv will print the compile date of the kernel currently running.
Modules:
Modules are used to reduce the amount of memory used to hold the kernel. There is a slight penalty for the time taken to load and unload the module.If the code is required for general operation of the kernel or is needed often or required by the boot process, it is best to compile it into the kernel and it should NOT be compiled as a module.
Module commands:
The kernel will use the modprobe utility to determine if the module is
compatible with the kernel.
The program used is specified by a proc file:
cat /proc/sys/kernel/modprobe
/sbin/modprobe - Result of cat operation
Modules are loaded by init scripts which call insmod/rmmod to load/unload modules.
Command | Description |
---|---|
lsmod | List loaded modules
[ /root]# lsmod Module Size Used by 3c59x 19716 0 (autoclean) (unused) autofs 9120 1 (autoclean) emu10k1 45688 1 soundcore 2596 4 [emu10k1] aic7xxx 137400 3 |
insmod | Inserts a module into the active kernel |
rmmod | Remove a loaded module. Just specify the module name. No ".o" or path necessary. |
modprobe | High level handling of loadable modules.
Loads module and dependencies. |
depmod | Creates dependencies file for a module (used by modprobe) |
modinfo | Display information about a kernel module |
List of modules held in /etc/modules.conf (or for older systems: /etc/conf.modules)
Sample:
alias scsi_hostadapter aic7xxx
alias eth0 eepro100
alias eth1 eepro100
alias parport_lowlevel parport_pc
alias sound-slot-0 emu10k1
alias char-major-81 bttv
Other useful commands:
- Display system status gathered from /proc: /usr/bin/procinfo -a
- List Processor type: cat /proc/cpuinfo
- List devices:
- cat /proc/devices
- /sbin/lsmod
- List IO ports (device address used drivers): cat /proc/ioports
- List DMA channels: cat /proc/dma
- View interrupts used by the system:
- cat /proc/interrupts
- procinfo -i
IRQ Description 0 Timer channel 0 1 Keyboard 2 Cascade for controller 2 (which controls IRQ 8-15) 3 Serial Port 2 (COM2) 4 Serial Port 1 (COM1) 5 Parallel Port 2 (LPT2) 6 Floppy Diskette Controller 7 Parallel Port 1 (Printer LPT1) 8 Real-time Clock 9 Redirected to IRQ2 10 -- 11 -- 12 PS/2 Mouse 13 Math Co-processor 14 Hard Disk Controller (IDE) 15 --
Also see file:- /etc/isapnp.gone for a list of protected IRQ's which will not be allocated to ISA Plug and Play device.
- Command: pnpdump
This command scans ISA PnP cards and can be used to generate /etc/isapnp.conf
pnpdump > /etc/isapnp.conf - Command: isapnp -C /etc/isapnp.conf
Uses file /etc/isapnp.gone and /proc files to avoid IRQ's already allocated and to determine safe IRQ's. - Plug and Play ISA Specification
- Create devices. Run script: /dev/MAKEDEV
- Display kernel version in current use: /proc/version
- Display boot messages: dmesg
- Display sound driver status: cat /dev/sndstat
[Potential Pitfall]: On platforms without much disk space (such as embedded platforms), a kernel compile can potentially fill up your drive potentially terminating the compile prematurely.
This is most often a result of the GNU C compiler cache hogging up as much as 0.5Gb to 1Gb of disk space in $HOME/.ccache/.
The GNU C compiler generates and uses this cache to speed up the compile.
This can be disabled by setting the following environment variable CCACHE_DISABLE by employing the command: export CCACHE_DISABLE=
The entire cache can be deleted and is no longer required once the environment variable has been set: rm -Rf $HOME/.ccache
In no way will the compiled results be any different but it will save allot of disk space.
Linux kernel naming convention: Version.MajorRelease.MinorRelease.PatchNumber
(In Linux kernel vernacular: KernelVersion.PatchLevel.Sublevel.Extraversion)
Patch File Description: The patch files are file "diffs" (differences from the stated source file) which can be applied to create a newer updated file.
Downloading kernel source and applying a patch:
- Obtain kernel source: ftp://ftp.us.kernel.org/pub/linux/kernel/v2.X/linux-2.X.X.tar.gz
Uncompress: tar xzf linux-2.X.X.tar.gz
(Or: tar xjf linux-2.X.X.tar.bz2)
This creates the local directory: linux/...
OR
use the source from the kernel-source RPM package released by your distribution. Using the source released by your distribution will include patches and source code needed to support all the features of your distribution. Of greatest importance is the default ".config" file which describes the features used by your distribution. i.e. Red Hat / Fedora Core use the ext3 file system, devfs and turn on many features by default. - Download patch:
- Download patch: ftp://ftp.us.kernel.org/pub/linux/kernel/v2.X/patch-2.X.X
- Apply patch:
Apply patches from default source directory /usr/src/linux (linked to actual source directory):
- gzip -cd ../KernelPatches/patch-file.gz | patch -p0
- bzip2 -dc ../KernelPatches/patch-file.bz2 | patch -p0
- If patch file resides in directory /usr/src/linux:
zcat patch-file.gz | patch -p0
(Script located in directory linked to kernel source)- Place all patches in /usr/src/KernelPatches/ (.gz or .bz2)
- Apply all the patches to kernel (i.e. 2.6.11.4)
/usr/src/linux/scripts/patch-kernel linux KernelPatches
(Execute from directory /usr/src) - Or apply all the patches to kernel to bring it up to specified level.
/usr/src/linux/scripts/patch-kernel linux KernelPatches 2.6.11.7
(patch-kernel kernel-source-directory-to-act-onhighest-version to include)
- Set-up directory links:
- mv linux linux-2.X.XX (only required on some older RPMs)
- ln -s linux-2.X.XX linux
- /usr/include/asm should point to /usr/src/linux/include/asm-i386
- /usr/include/linux should point to /usr/src/linux/include/linux
- /usr/include/scsi should point to /usr/src/linux/include/scsi
Example: Applying the devfs patch to kernel 2.2.20
- Download kernel: http://www.kernel.org/pub/linux/kernel/v2.2/linux-2.2.20.tar.gz
- mv linux-2.2.20.tar.gz /usr/src
- cd /usr/src/
- Remove previous source: rm linux (May use rm -Rf linux/)
- Unload to subdirectory linux/: tar xzf linux-2.2.20.tar.gz
- Download patch: http://www.atnf.csiro.au/~rgooch/linux/kernel-patches.html
- Download patch file devfs-patch-current to /usr/src/
- cd /usr/src
- Test: patch -p0 --dry-run < devfs-patch-current
- Apply patch: patch -p0 < devfs-patch-current
- Now go through regular kernel build as described above. (From the beginning) make distclean; make mrproper
- The configuration has been altered by the patch thus make xconfig
will have options not seen in the original 2.2.20 kernel.
- "Code maturity level" option
- Select "development and/or incomplete code/drivers"
This will provide lots of options for kernel. - Go to Filesystems - check /dev file support "Experimental"
This feature was introduced by the patch. If it is not selected the devfs feature introduced by the patch will not be included.
Patch command options:
Option | Description |
---|---|
-DMACRO-NAME | Define pre-processor macro |
-c | Use diff file format. |
-d directory | Change to directory. |
--dry-run | Don't modify file but output results. |
-s | Silent mode. |
-f | Force change. |
-n | Using patch file which is a regular diff file. |
-R | Reverse or undo patch. (i.e. if patch rejected.) |
-p0 | Do not strip directory path prefix. Path intact. i.e. /usr/src/linux/... |
-p1 | Strip first "/" in directory path. i.e. usr/src/linux/... |
-p2 | Strip first directory in path. i.e. src/linux/... |
-p# | Directory path "/" prefixes to strip (# - 1). |
View patch file: zcat patch-file.gz | more
- Intel P4 processor with hyperthreading will act as two CPU's. Use an SMP Linux kernel to get maximum performance.
- RHEL6: Linux Performance Tuning Guide
- System Tuning - Disk, SCSI, TCP, Threads etc.
- Raid Optimization for Database Operations
- Red Hat Linux General Optimization
- IBM Redbook: Performance tuning - [PDF]
- Linux Kernel Configuration Help Texts
- hardware configuration utilities
- LDP: Kernel HOWTO
- Kernel.org Linux Kernel
- Uniform Driver Interface - SCO released UDI to the Open Source World to provide an environment for portable driver code.
- Linux ECC - ECC memory support
- IRQTUNE -- A Linux IRQ Priority Optimizer - changes the IRQ priority of devices to allow devices that require high priority and fast service (e.g. serial ports, modems) to have it.
- Browse the kernel source - The Linux Cross-Reference project. (LXR)
- BitWizard - Device Driver consultants
- Scalability Limits and Issues - Dan Kegel
- Linux Scalability Project
- Understanding Virtual Memory (pdf)