This is an old revision of the document!


System configuration

How do I get an out-of-the-box working Linux audio workstation?

How do I build a real-time audio workstation on Linux?

Preliminary

The modifications discussed in the following sections have been tested thoroughly and a number of these modifications have been implemented in specialized distros like KXStudio and AVLinux. This doesn't guarantee though that these modifications work for all set-ups and these modifications shouldn't be used as is but should be adapted if possible to match your own set-up. It is also advised to try the modifications step by step, so do not implement them all at once. This way you can verify if each modification does what it claims to do. By implementing the modifications a step at a time it is easier to track down which modification is causing any unwanted behavior.

A good habit is to always make backups of any files you're about to modify. Take /etc/fstab for instance. Before adding the noatime mount parameter you might want to make a backup of your existing and working fstab file:

sudo cp /etc/fstab /etc/fstab.orig

This way you can restore the original file in the case you made such a typo that the system has become unbootable. The procedure of recovering an unbootable system itself is beyond the scope of this Wiki but there are plenty of resources that explain this in detail.

QuickScan

The realTimeConfigQuickScan script automatically analyzes your current configuration. This script will make configuration suggestions based on your current setup and link to the relevant sections on this page for background information.

To get the script install Git first and then pull in the script:

git clone git://github.com/raboof/realtimeconfigquickscan.git
cd realtimeconfigquickscan
perl ./realTimeConfigQuickScan.pl

Discuss this script at the LinuxMusicians Forum

The kernel

Note about kernels

  • Kernels >= 2.6.31 seem to work pretty good without RT patch, also for real-time pro audio usage. It's not strictly necessary anymore to install a real-time ('rt') kernel to get good results. Although the best results are still expected when using a real-time kernel. Try it, test it and decide for yourself.
  • In the pre-2.6.39 kernel era rt kernels were indeed necessary in some cases where sound devices were sharing IRQ's with other peripherals. With the rt kernel and the rtirq script you could prioritize IRQ threads but since 2.6.39 it is possible to use the rtirq kernel with a generic kernel and the threadirqs kernel option.
  • Many distro kernels, and many third-party-provided real-time kernels, are configured without the 1000 Hz setting. This is not significant for recording, but it can be crucial issue for live MIDI work. If your application is live MIDI and using ALSA MIDI, you may well want real-time, but you certainly need 1000 Hz.

Status of your kernel

Run uname -a to find out which kernel you're running:

Linux xps13 4.8.15-core2-rt10 #1 SMP PREEMPT RT Sun May 28 10:43:44 CEST 2017 x86_64 x86_64 x86_64 GNU/Linux

4.8.15-core2-rt10 is the version number of the currently running kernel. The configuration file with which this kernel was created can be found in /boot/config-4.8.15-core2-rt10. Open it, and check if it has the following characteristics:

CONFIG_HZ_1000=y
CONFIG_HZ=1000
CONFIG_PREEMPT_RT_FULL=y
CONFIG_PREEMPT=y

The “RT” is real-time, and the above reflects a real-time kernel. The “1000” reflects the 1000 Hz setting.

If you can't find the config file for your kernel there is also the possibility to extract the config from /proc/config.gz if your kernel supports it (CONFIG_IKCONFIG_PROC=y):

cat /proc/config.gz | gunzip > config-`uname -r`

The Quickscan is also good for checking if you have a RT kernel and if the kernel timer setting is adequate.

Do I really need a real-time kernel?

As of kernel 2.6.39 one the core parts of the real-time patchset, threaded IRQs, has been integrated into the main kernel. With the kernel boot option threadirqs standard kernels create threaded IRQs, a feature that was reserved to kernels < 2.6.39 patched with the real-time patch-set. Read more about threaded IRQs here. The integration of this feature into the main kernel basically nullifies the need to use a real-time kernel in most of the cases. But a real-time kernel can still be very useful on a Linux audio platform. If you need the lowest latencies possible and tight timing (for live MIDI applications for instance) a real-time kernel might still be your best resort. But if you're mainly recording, a standard kernel or an optimized kernel (like the low-latency kernel Ubuntu offers) will suffice probably.

Using the threadirqs kernel option

This is only needed for so-called generic kernels, ie. standard kernels that are not tweaked for lowlatency performance. You can check if your kernel already includes this option with the following command:

$ grep -e "CONFIG_IRQ_FORCED_THREADING=y" -e "CONFIG_PREEMPT=y" /boot/config-`uname -r`

If it returns CONFIG_IRQ_FORCED_THREADING=y and CONFIG_PREEMPT=y then your kernel is using IRQ threading and you don't have to proceed with the following steps. If it only returns CONFIG_IRQ_FORCED_THREADING=y you can add the threadirqs boot option as described below. If the command returns nothing you have a kernel that cannot use threaded IRQs.

Open /etc/default/grub with your favorite editor as root. Look for the line that starts with GRUB_CMDLINE_LINUX_DEFAULT and add threadirqs to the list of options:

GRUB_CMDLINE_LINUX_DEFAULT="quiet splash threadirqs"

Save the file and update the grub configuration:

sudo update-grub

Now reboot and you should have threaded IRQs.

Installing a real-time kernel

This section describes installing a real-time kernel on an existing Linux distribution. Some distributions contain packages you can install to get a real-time kernel, for others you'll need to compile it yourself. You can recognize real-time kernel versions by the added '-rt' or '-realtime' parts of the kernel version number.

Ubuntu

Ubuntu Studio 12.04 comes with a 1000 Hz non-real-time low-latency kernel. Unfortunately there are no packaged real-time kernels available but using the kernel build method described in this Wiki it should be fairly easy to build your own real-time kernel1).

Debian

Since Debian Wheezy an RT variant is also available in the stock repositories; unfortunately its system timer is set to 250 Hz instead of 1000 Hz, so you'll probably want to recompile your own kernel anyway, especially if you use ALSA MIDI. These kernels incorporate Pengutronix patches (see DebianMultimedia Wiki). The 64 Studio real-time kernels are in the backports apt repository, so for Debian Squeeze you will have to add the following line to your apt sources list:

deb http://apt.64studio.com/backports squeeze-backports main
Arch

Arch has rt kernels available in AUR.

Gentoo

RT kernels are vailable in the pro-audio overlay (http://proaudio.tuxfamily.org/wiki/index.php?title=Main_Page).

Build your own real-time kernel

If your distribution isn't any help, you can compile and install a real-time kernel manually. The examples assume building a RT kernel based on the 3.2 kernel version but they should generally apply fo other kernel versions too provided that there is a RT patchset available for the used kernel.

Install the necessary packages, on Debian and Ubuntu this is:

sudo apt-get install kernel-package fakeroot build-essential

Download the kernel sources and the RT patchset (this example uses the 3.18 branch, you're free to use another kernel branch for which there is a RT patch set available):

mkdir -p ~/tmp/linux-rt
cd ~/tmp/linux-rt
wget -c https://www.kernel.org/pub/linux/kernel/v3.x/linux-3.18.7.tar.xz
wget -c https://www.kernel.org/pub/linux/kernel/projects/rt/3.18/patch-3.18.7-rt1.patch.xz

Extract the kernel sources and patch them with the RT patchset:

tar xJvf linux-3.18.7.tar.xz
cd linux-3.18.7
xzcat ../patch-3.18.7-rt1.patch.xz | patch -p1

Now you need to configure the kernel. The easiest way is to use an existing kernel config, those configs can be found in the /boot/ directory of your system. You can copy an appropriate kernel config to your current working directory and use that config as a starting point:

cp /boot/config-`uname -r` .config

This will copy the config of the kernel you're currently using. You might want to consider using a config of a kernel already optimized for audio work, like the Ubuntu -lowlatency kernel config. The next step is to create a new config with full preemption enabled from the copied config:

make oldconfig

You will get a lot of prompts but you can leave everything at its default value except for the prompt which Preemption Model you'd like to use. Select option 5 there (Fully Preemptible Kernel):

Preemption Model
> 1. No Forced Preemption (Server) (PREEMPT_NONE)
  2. Voluntary Kernel Preemption (Desktop) (PREEMPT_VOLUNTARY)
  3. Preemptible Kernel (Low-Latency Desktop) (PREEMPT__LL) (NEW)
  4. Preemptible Kernel (Basic RT) (PREEMPT_RTB) (NEW)
  5. Fully Preemptible Kernel (RT) (PREEMPT_RT_FULL) (NEW)
choice[1-5]: 5 <Enter>

You can use one of these commands to edit the configuration:

make config

if you want a text interface, or

make menuconfig

for a nice ncurses text interface (you need to install libncurses5-dev if you're on Debian or Ubuntu), or

make nconfig

for another ncurses interface with better keybindings, or

make xconfig

for a qt GUI interface (you need to install libqt4-dev if you're on Debian or Ubuntu).

Make sure to check the “Timer Frequency: 1000 Hz” options under the “Processor type and features” group.

When your configuration is done, you can build and install the kernel and its modules (distro-agnostic way):

make
make install
make modules_install

On Debian or Ubuntu the building step can be done like this:

make -j `nproc` LOCALVERSION= deb-pkg

Or if you prefer make-kpkg:

CONCURRENCY_LEVEL=`nproc` LOCALVERSION= fakeroot make-kpkg --initrd kernel_image kernel_headers

You can also use `getconf _NPROCESSORS_ONLN` instead of `nproc`.

When building is complete and the packages have been created you can install them (this will also update the boot loader menu):

cd ..
sudo dpkg -i linux-{headers,image}-3.18.7-rt1_*.deb

If you built and installed the kernel the distro-agnostic way you will have to update the bootloader yourself:

sudo update-grub
Optional bits

You may want to disable the debug features; this will save compile space. You can do it in the config interface, or from the shell with this command:

scripts/config --disable DEBUG_INFO

You may want to set the “Local version” (in the “General setup” group) to your name or -custom-rt or whatever: it will be appended to the kernel name so you'll easily recognize it as a custom build, and you won't risk overwriting some working kernel with the same name. You can achieve the same result by editing the file localversion-rt and append a local version name to the version number of the RT patch set.

You can do without an initrd if you build into the kernel (not as a module, i.e. checked in the config as “*” and not “m”) everything you need to boot (e.g. ext4, chipset drivers, keyboard drivers, etc.). If you need it, build it with:

mkinitrd -k kernel-<kernel version> -i initrd-<kernel version>

or follow your distro guide.

If you need to know which drivers you need for your hardware, use the form at http://kmuto.jp/debian/hcl/ (it's distro-agnostic, even if it says “Debian” everywhere).

Build your own real-time kernel on Debian Wheezy or later

Since Debian Wheezy, the linux-source-3.2 package already installs the RT patch (you don't need to download it from upstream), but you'll have to apply it manually.

(Instructions taken from http://kernel-handbook.alioth.debian.org/ch-common-tasks.html).

Install the kernel source package:

sudo aptitude install linux-source

This installs a linux tarball and a bzipped RT patch in /usr/src/ . Copy them to wherever you want to build (you'll need some GB of free space), untar/bunzip them and cd into the linux source directory.

Apply the RT patch:

patch -p1 -i ../linux-patch-3.x-rt.patch

Make sure to install the needed tools for the build:

sudo apt-get build-dep linux

Now you must configure the kernel. (This step is distro-agnostic, you can follow the istructions in the previous paragraph).

Then you can compile…

make clean
make deb-pkg

… and install the new kernel

sudo dpkg -i ../linux-image-3.x.yy_foo.deb

Disabling resource-intensive daemons, services and processes

You should check if the following daemons, services or processes are running, and if so, consider stopping them and/or resort to alternatives:

  • powersaved
  • kpowersave
  • upowersave ?
  • CPU frequency scaling daemons that scale the frequency of the CPU depending on the CPU load could cause xruns in some cases. More recent versions of Jack1 (>= 0.118.0) are suffering less from xruns or run xrun free with a CPU frequency scaling daemon enabled that's set to a scaling governor like ondemand. A specialized CPU scaling daemon is in the works that depends on the DSP-load instead of the CPU load: jackfreqd
  • Debian/Ubuntu: a default install will have the apt-xapian-index package installed that enables the Quick Search option in the Synaptic package manager. In order for this Quick Search functionality to work properly it needs to regularly update its database with the help of a cronjob. Especially on less powerful machines this cronjob could cause major xruns. Consider uninstalling the apt-xapian-index package.
  • Desktop Environments (DE): especially on less powerful machines it is recommended to use a lighter DE than Gnome, KDE or Unity. Possible alternatives include LXDE, XFCE or IceWM. Another option is to only use a lightweight Window Manager (WM) like Openbox or Fluxbox.
  • Compositing: same recommendation as for the DE you use. Try to avoid using a compositing manager like Compiz to lower the chance of running into xruns.
  • Gnome NetworkManager and WiFi: NetworkManager keeps scanning for new wireless networks in the background and this might cause xruns. The best option is to not use a wireless network manager in a low-latency real-time audio environment. If you really can't do without WiFi you probably won't have another option than to use wpa_supplicant.
  • NTP: there are reports of NTP interfering with FFADO, so if you're using a FireWire audio device make sure you disable the ntpd service and any cronjob or ifup script that tries to sync your local clock with an NTP server. This issue is resolved in the 2.1.0 release of FFADO.

CPU frequency scaling

If your CPU supports frequency scaling and the CPU frequency scaling governor is set to ondemand (which is the default on a lot of distros) you could run into xruns. The ondemand governor scales the frequency according to the CPU load, the more the load, the higher the frequency. But this is happening independently from the DSP load on your system so it could happen that the DSP load suddenly rises for instance, demanding more CPU power, and that the scaling daemon kicks in too late, resulting in xruns because the DSP load maxes out. A solution would be to use a CPU frequency scaling daemon that scales the frequency according to the DSP load on your system like jackfreqd or to simply disable CPU frequency scaling altogether. The latter can be achieved by setting the scaling governor to performance.

To check which governor is active:

cat /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor

Setting the governor to performance:

echo -n performance | sudo tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor

You could also add a line to your /etc/rc.local file for instance to set the governor to performance at boot time:

echo -n performance > /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor

Ubuntu

On Ubuntu systems the command in your /etc/rc.local file only works if you disable the ondemand service:

sudo update-rc.d ondemand disable

Another option would be to modify the ondemand init script and rename it to performance:

sudo sed -i 's/ondemand/performance/g' /etc/init.d/ondemand
sudo update-rc.d ondemand disable
sudo cp /etc/init.d/ondemand /etc/init.d/performance
sudo update-rc.d performance defaults

Debian

On Debian you can control the scaling governor with the cpufreq-set utility which is part of the cpufrequtils package:

sudo cpufreq-set -r -g performance

On installation the package also installs an init script in /etc/init.d/cpufrequtils and a configuration file in /etc/default/cpufrequtils. To have the governor always set to performance make sure the configuration file looks as follows:

ENABLE="true"
GOVERNOR="performance"
MAX_SPEED="0"
MIN_SPEED="0" 

BIOS

Most BIOSes allow to disable CPU frequency scaling altogether and as a result your CPU will run at its maximum clock rate. For Intel CPUs you should look for a SpeedStep option. Its AMD equivalents are either Cool'n'Quiet or PowerNow!.

If you really need to prevent the CPU from running hot (to prevent shutdowns or overheating) or if you prefer your PC's to consume less power then you can also adjust some BIOS settings (if your BIOS has them):

  • Disable PCI and/or PCIe power-management in the BIOS
  • Disable EIST and C1E halt-states

With these settings disabled the ondemand governor should work without causing any xruns.

Disabling problematical services and other hidden bits

Types of things to disable

There are several different types of things which can be helpful to disable. But within all of the types, are likely to be items which you will need for a working machine! So do be careful.

  • System services using the SysV / rc.d / init.d standards.
  • System services using the newer 'upstart' standard.
  • DBus-initiated processes.
  • Hardware support modules.
  • GUI login session startup items.

System services using SysV / rc.d / init.d.

Many distros still use this exclusively. Command-line tools for disabling these are update-rc.d and chkconfig. On most distros which use these, you'll need the concept of the “runlevel”; a runlevel is a level at which the system can be run, and each runlevel has its own set of services. The older standard was to have separate runlevels for single-user mode for debugging, text-mode only, X, etc., but there is a lot of variation on this now, rather distro-specific.

System services using upstart.

Upstart is quite new, fast, and commended by much of the Linux devel community. Ubuntu 12.04 uses a combination of SysV and upstart, and its community is undergoing a migration towards upstart only. To disable services controlled by upstart, go to the folder /etc/init (not /etc/init.d, that's SysV), find the name of the config file corresponding to an item you want disabled (e.g., bluetooth.conf), create a corresponding file of extension .override (in this case, bluetooth.override). The .override file needs to contain just the one line “manual”. Reboot to implement the change.

DBus-controlled services

DBus is a solid standard used for a wide variety of services which need to be started automatically, but on demand only, in the background, not specifically at boot. One set of DBus services used extensively in Ubuntu is 'gvfs', the virtual filesystem for the Gnome desktop; if these are permitted to run, polling of hardware including the USB bus will occur, which is likely to interfere with USB MIDI interfaces, eat CPU power, and cause xruns in general. To disable gvfs, go as root to /usr/share/dbus-1/services , rename gvfs-daemon.service to gvfs-daemon.service.disabled, and rename gvfs-metadata.service to gvfs-metadata.service.disabled.

Hardware support modules

These are modules in the Linux kernel which support various hardware. All of them will reserve resources. If, for instance, you have an HDMI audio port which you don't need, you can blacklist it to disable it. However, you should check the IRQ usage first (see related section of this document), because this might be the more vital change to make, if you do in fact have a conflict.

Example script for needed disables

There are quite some processes and drivers running that you might not need when making music. Below an extremely simple example of a Bash script to disable all unneeded things. Note that it includes a solution for IRQ conflict, documented further down.

#!/bin/sh

modprobe -r ppdev # I don't have a parallel port
modprobe -r lp # I don't use a printer when making music
modprobe -r uvcvideo # I don't use a webcam when making music
modprobe -r videodev # Ditto
modprobe -r ath9k # Wireless driver
modprobe -r r8169 # NIC driver
modprobe -r btusb # Bluetooth USB

/etc/init.d/bluetooth stop & # Stop Bluetooth related processes
/etc/init.d/cups stop & # Stop cups, I don't use a printer when making music
/etc/init.d/networking stop & # Stop networking, the internet is a very distracting factor when making music
/etc/init.d/network-manager stop & # Ditto

killall modem-manager # I don't have a modem
killall wpa_supplicant # See the network stuff above

modprobe snd-hrtimer # Load the ALSA high res timer for my MIDI stuff

echo -n performance > /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor # Set the CPU scaling governor to performance

echo -n "0000:00:13.0" > /sys/bus/pci/drivers/ohci_hcd/unbind # Unbind a USB port, this one interferes with my onboard soundcard

# The following section is for kernels < 2.6.39. Kernels > 2.6.39 do not have a tasklet daemon.
TASKLETPR=76 # Define a variable TASKLETPR and set it to 76

ps -eLo pid,cmd | grep [t]asklet | awk '{ system("chrt -f -p '$TASKLETPR' " $1)}' # Do some greppy awky stuff and set the prio of the tasklet daemon

You could run this script once you're logged in (it needs root privileges though), run it from /etc/rc.local or even as an init script. A suggestion would be to run it only when a real-time kernel is loaded for example:

if [ $(uname -r | cut -d "-" -f 3) = "realtime" ]
then /path/to/the/above/script
fi

The modules and processes differ of course on a per machine basis so first figure out which modules are loaded with lsmod for instance and then decide which ones you really need. Same for any processes running on your machine, ps -ef or even pstree -p could be helpful when looking for unneeded processes.

Solve IRQ conflicts by unbinding devices

The above script also contains a line to unbind a rogue device that sits on the same IRQ as the sound card. Unbinding a device is basically a last resort if a device is interfering in such a way that all other possibilities, like using rtirq, fail to improve the performance of the sound card. In this particular case the thread created for the USB port started consuming so much CPU that at a certain moment xruns were all over the place. In order to prevent this from happening unbinding the USB driver from the physical USB port worked out well.

This is the situation on the affected machine:

$ cat /proc/interrupts | grep 16
16:        923   IO-APIC-fasteoi   ohci_hcd:usb2, hda_intel

So usb2 shares IRQ 16 with the HDA Intel soundcard. In order to be able to unbind the USB driver from usb2 we need to find the bus ID of usb2:

$ tree /sys/bus/usb/drivers/usb/
/sys/bus/usb/drivers/usb/
|-- 1-4 -> ../../../../devices/pci0000:00/0000:00:13.5/usb1/1-4
|-- bind
|-- uevent
|-- unbind
|-- usb1 -> ../../../../devices/pci0000:00/0000:00:13.5/usb1
|-- usb2 -> ../../../../devices/pci0000:00/0000:00:13.0/usb2
|-- usb3 -> ../../../../devices/pci0000:00/0000:00:13.1/usb3
|-- usb4 -> ../../../../devices/pci0000:00/0000:00:13.3/usb4
`-- usb5 -> ../../../../devices/pci0000:00/0000:00:13.4/usb5

6 directories, 3 files

So the ID of usb2 is 0000:00:13.0. If we write that to the unbind file in the /sys/bus/pci/drivers/ohci_hcd/ directory usb2 should get unbound from the USB driver:

# echo -n "0000:00:13.0" > /sys/bus/pci/drivers/ohci_hcd/unbind

Finding the right unbind file might be a bit trial and error, but /sys/bus/pci/drivers/ohci_hcd/ contains the right unbind file for USB1.1 root hubs such as usb2. Now we can check with the first tree command (or with lsusb) what the current status is:

$ tree /sys/bus/usb/drivers/usb/
/sys/bus/usb/drivers/usb/
|-- 1-4 -> ../../../../devices/pci0000:00/0000:00:13.5/usb1/1-4
|-- bind
|-- uevent
|-- unbind
|-- usb1 -> ../../../../devices/pci0000:00/0000:00:13.5/usb1
|-- usb3 -> ../../../../devices/pci0000:00/0000:00:13.1/usb3
|-- usb4 -> ../../../../devices/pci0000:00/0000:00:13.3/usb4
`-- usb5 -> ../../../../devices/pci0000:00/0000:00:13.4/usb5

5 directories, 3 files

As you can see, usb2 is gone, it doesn't show up with lsusb anymore either:

$ lsusb
Bus 005 Device 001: ID 1d6b:0001 Linux Foundation 1.1 root hub
Bus 004 Device 001: ID 1d6b:0001 Linux Foundation 1.1 root hub
Bus 003 Device 001: ID 1d6b:0001 Linux Foundation 1.1 root hub
Bus 001 Device 002: ID 04f2:b175 Chicony Electronics Co., Ltd 
Bus 001 Device 001: ID 1d6b:0002 Linux Foundation 2.0 root hub

usb2 used to correspond with Bus 002. That one seems to be gone. A final check:

$ cat /proc/interrupts | grep 16
 16:        924   IO-APIC-fasteoi   hda_intel

Great, the sound card is now sitting on its own on IRQ 16 and runs in a way more stable manner than before when it still shared an IRQ with the USB port. Major drawback of this method of course is that you loose an USB port.

More information: http://lwn.net/Articles/143397/

Filesystems

For audio use, it is desirable to use a filesystem that favours few big files over many small files and low-latency over long-term thoughput. It is said that ReiserFS and fuseblk are bad choices from this perspective, while ext3 and ext4 are good. ext4 has had some performance issues in the past but those were not directly related to processing audio and in the meanwhile most issues have been ironed out.

In any case, it is advisable to mount ext filesystems with the noatime option enabled. This step is verified by the QuickScan, but could use more information about which filesystems are suitable.

SSD

Most distributions currently do not auto-detect and optimize your filesystem settings for SSD drives. Here is a guide on how to optimize your SSD for use under Ubuntu-based distros. There is also a PDF document available on elinux.org on how to tune your filesystem for use with a SSD drive: http://elinux.org/images/b/b6/EMMC-SSD_File_System_Tuning_Methodology_v1.0.pdf

Encryption

It won't affect the quality of the audio, but it will change how many tracks the disk can support and because encryption requires CPU, it will affect how much signal processing you can do all of which is a long-winded way of saying “don't do it

tmpfs

Both Jack1 and Jack2 now use /dev/shm/ so mounting /tmp/ to tmpfs is not necessary anymore.

ext3/ext4 tuning options

noatime

Setting the noatime parameter in your /etc/fstab file reduces the amount of disk I/O (the inode access times are not updated each time a file is read) which could improve the overall performance of your system.

/dev/sdax / ext4 noatime,errors=remount-ro 0 1

Most modern distributions use either atime or relatime. See the mount manpage for more information on these parameters.

The noatime parameter implies nodiratime so when using noatime there's no need to set nodiratime.2)

nobarrier

According to this thread High latency with ext4 / jbd2 kernel thread on the ext4 mailing list, the nobarrier parameter, with an ext4 filesystem, can reduce the amount of work made by the background kernel thread jbd2. However, it is said to be an unsafe parameter, making your filesystem less resistant to power failures. Use it with caution.

/dev/sdax / ext4 nobarrier,errors=remount-ro 0 1

Ext3 filesystem doesn't enable barrier by default, ext4 does. See the mount manpage for more information on these parameters.

data=writeback
This option specifies the following journaling mode for file data: data may be written into the main file system after its metadata has been committed to the journal. This is rumored to be the highest-throughput option. It guarantees internal file system integrity, however it can allow old data to appear in files after a crash and journal recovery.3)

You can't use this parameter in /etc/fstab for the root partition, you will have to set it as a kernel parameter. So when using Grub you can add the following parameter to the GRUB_CMDLINE_LINUX_DEFAULT line in /etc/default/grub:

rootflags=data=writeback

Use this parameter with caution.

limits.conf/audio.conf

It is advisable to set up your /etc/security/limits.conf or /etc/security/limits.d/audio.conf file properly, for example:

 @audio - rtprio 90       # maximum realtime priority
 @audio - memlock unlimited  # maximum locked-in-memory address space (KB)

Setting memlock to unlimited shouldn't be strictly necessary since most apps would work fine with lower values (such as 500000).
However, certain applications have been reported to complain or even to crash with lower values than unlimited.
On the other hand, giving unlimited capability of locking memory can cause buggy applications to temporarily freeze the whole system.
See http://www.linuxmusicians.com/viewtopic.php?f=10&t=2193 for more details.

You could also allow the audio group to renice processes with the help of the limits.conf file, but since nice uses SCHED_OTHER it basically does nothing to increase the performance of a low-latency real-time audio environment that relies on SCHED_FIFO/SCHED_RR. Consult the sched_setscheduler manpage for more info on this subject.

See also: http://linuxmusicians.com/viewtopic.php?f=27&t=392

If you do not use PAM (which is rare), you might want to use set_rlimits instead

sysctl.conf

Suggested settings for /etc/sysctl.conf:

 vm.swappiness = 10

This setting changes the so-called swappiness of your system, or in other words, the moment when your system starts to use its swap partition. You can check the current value with cat /proc/sys/vm/swappiness, in most cases it is set to 60. This is too high, it will make your system start using swap too fast which might influence the overall performance.

There are references on the net on adjusting the fs.inotify.max_user_watches value also for enhanced performance. But it remains very unclear where these references come from and if adjusting this value actually does anything at all. The max_user_watches parameter sets the maximum number of files your system can monitor with inotify (which is part of the kernel) for changes. Setting this parameter too low results in inotify failing. Setting it too high can make inotify needlessly consume memory. Best is to not touch the default settings as setting this parameter is unrelated to performance in an audio context.

audio group

It is generally good practice to have an 'audio' group, and add any users that should be allowed to perform audio tasks to this group. This prevents some interference from non-audio-processes with audio tasks. To verify you're in the 'audio' group, run the groups command. Remember, after adding yourself to new groups, you need to log out and back in again. Be careful when adding an 'audio' group to your system: most systems come with a pre-configured 'audio' group, and do not warn you when you add another group with the same name, leading to much confusion.

This step is verified by the Quickscan.

Timers

The configuration of hardware or software timers only matters for ALSA MIDI performance. JACK MIDI, JACK's MIDI backend drivers (like the alsarawmidi driver) and tools like a2jmidid do not care about the kernel clock frequency. When using applications that rely on ALSA MIDI (like Qtractor, SEQ24 or Rosegarden) a high resolution timer might reduce MIDI jitter and improve MIDI performance.

Hardware timers

:!: Using hardware timers to achieve more accurate MIDI timing is questionable. Applications that use ALSA MIDI (via either the Sequencer or the Raw MIDI Interface) may be affected by using hardware timers but it might be better to rely on a high resolution software-based timer like the snd-hrtimer kernel module.

Some MIDI sequencers or DAWs that rely on ALSA MIDI might benefit from being able to use hardware timers like the real-time clock (/dev/rtc) or the High Precision Event Timer (/dev/hpet). When using a hardware timer make sure the 'audio' group has read permissions on it.

A simple chgrp might not be persistent across reboots - to make the change last, create a new 40-timer-permissions.rules file in /etc/udev/rules.d with the following lines:

KERNEL=="rtc0", GROUP="audio"
KERNEL=="hpet", GROUP="audio"

It is also advisable to enable users of the audio group to use these timers at higher frequencies since the defaults are too low:

$ cat /proc/sys/dev/hpet/max-user-freq 
64
$ cat /sys/class/rtc/rtc0/max_user_freq 
64

Unfortunately there's no consensus on what values to use. Most sources mention a minimal value of at least 1024 or even 2048 and a maximal setting of 8192. It might be wisest to use the values mentioned in the Rosegarden Wiki as Rosegarden is an application that relies on accurate max user freq settings. The max user freq setting for hpet can be set in /etc/sysctl.conf as it lives in /proc and the setting for rtc0 can be set with an echo command at boot-time with either a start-up script or adding a line to /etc/rc.local.

/etc/sysctl.conf (or something like /etc/sysctl.d/60-max-user-freq.conf):

dev.hpet.max-user-freq=3072 

Start-up script or /etc/rc.local:

echo 3072 >/sys/class/rtc/rtc0/max_user_freq

And reboot. To activate these settings immediately:

sudo sysctl -p /etc/sysctl.d/60-max-user-freq.conf
echo -n 3072 | sudo tee /sys/class/rtc/rtc0/max_user_freq
sudo chmod 660 /dev/hpet /dev/rtc0
sudo chgrp audio /dev/hpet /dev/rtc0

See also: https://bugs.launchpad.net/ubuntu/+source/udev/+bug/306458

This step is verified by the QuickScan.

Software timers

:!: WIP

Priorities

For rtprio, higher values mean higher priority. You can list threads and their priorities with one of the following commands

$ ps axHo user,lwp,pid,rtprio,ni,command
$ ps -eLo pid,cls,rtprio,pri,nice,cmd | grep -i irq

Or if you have rtirq installed:

$ /etc/init.d/rtirq status

Most applications take care of running some threads with higher rtprio when possible themselves, like JACK with the real-time option set. Jack1's watchdog has a higher prio than jackd itself (+10) so choose a sensible priority value for Jack1 (70 for example, this will assign prio 80 to Jack1's watchdog thread), bearing in mind also that sound devices might get a priority in the 80-90 range if you use rtirq. Jack2 doesn't have a watchdog thread (neither does the current git version of Jack1) so you can give Jack2 an even higher prio.

You can also set rtprio and scheduling settings manually using the chrt command-line tool, which should come with most distributions nowadays, or use the rtirq init script.

See also: http://subversion.ffado.org/wiki/IrqPriorities

rtirq

rtirq is a script written by Rui Nuno Capela, the author of QjackCtl and Qtractor amongst others. This script allows to make use of the threaded IRQs as used by real-time kernels or kernels >= 2.6.39 with the threadirqs kernel option enabled. The term threaded IRQ is quite an abstract one but what it boils down to is that every peripheral gets an IRQ and the Linux kernel splits this IRQ in different parts of which the so-called bottom-halve is of concern in the context of the rtirq script. Bottom-halves are also known in some cases as softirqs which might make this matter a bit less abstract, 'hardirqs' (the top-halves) are the IRQs assigned by the system which can't be managed from within the OS as opposed to 'softirqs' (the bottom-halves) which can be managed.

These bottom-halves are managed through so-called tasklets that are part of the tasklet API. This functionality can be verified with the aforementioned ps -eLo pid,cls,rtprio,pri,nice,cmd | grep -i “irq” command. The process called sirq-tasklet (<= 2.6.38, kernels >= 2.6.39 do not have such a process simply because that specific part from the RT patchset has been merged into the main kernel) is the actual tasklet management tool that takes care of assigning the priorities to different bottom-halves, or actually the threads in which these bottom-halves run (hence the term threaded IRQs). The distinction between these three terms (softirq, bottom-halve and tasklet) is a bit vague but basically tasklets are a top and bottom halve implementation based on softirqs. If you'd like to know more on this issue the following links are a good place to start:

The rtirq tarball as available on the site of the author consists of two files, the script itself (rtirq.sh) and a configuration file (rtirq.conf). In most cases the rtirq.sh file is used as an init script (and hence stripped from the .sh extension) and rtirq.conf mostly ends up in /etc/default/ or /etc/sysconfig/, also stripped from its extension. Let's look at the contents of the configuration file:

#
# Copyright (c) 2004-2011 rncbc aka Rui Nuno Capela.
#
#   This program is free software; you can redistribute it and/or
#   modify it under the terms of the GNU General Public License
#   as published by the Free Software Foundation; either version 2
#   of the License, or (at your option) any later version.
#
#   This program is distributed in the hope that it will be useful,
#   but WITHOUT ANY WARRANTY; without even the implied warranty of
#   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
#   GNU General Public License for more details.
#
#   You should have received a copy of the GNU General Public License along
#   with this program; if not, write to the Free Software Foundation, Inc.,
#   51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
#
# /etc/sysconfig/rtirq
# /etc/default/rtirq
#
# Configuration for IRQ thread tunning,
# for realtime-preempt enabled kernels.
#
# This program is free software; you can redistribute it and/or modify it
# under the terms of the GNU General Public License version 2 or later.
#

# IRQ thread service names
# (space separated list, from higher to lower priority).
RTIRQ_NAME_LIST="rtc snd usb i8042"

# Highest priority.
RTIRQ_PRIO_HIGH=90

# Priority decrease step.
RTIRQ_PRIO_DECR=5

# Whether to reset all IRQ threads to SCHED_OTHER.
RTIRQ_RESET_ALL=0

# On kernel configurations that support it,
# which services should be NOT threaded 
# (space separated list).
RTIRQ_NON_THREADED="rtc snd"

# Process names which will be forced to the
# highest realtime priority range (99-91)
# (space separated list, from highest to lower priority).
# RTIRQ_HIGH_LIST="timer"

The RTIRQ_NAME_LIST variable contains a list of space separated service names of which the first entry gets the highest priority. The term service seems to refer to module names and sound device designations (so the output of lsmod and aplay -l respectively) and doesn't have to correspond to the full output, part of the output may suffice as the rtirq script does the matching itself.

The RTIRQ_PRIO_HIGH variable sets the highest priority a service can get assigned.

The RTIRQ_PRIO_DECR speaks for itself, here you set the number with which the priorities for each consequent service mentioned in the RTIRQ_NAME_LIST variable should be decreased.

The RTIRQ_RESET_ALL is a legacy variable and can best be left to default.

The RTIRQ_NON_THREADED variable is another legacy option, your kernel configuration has to support it and in almost all cases it doesn't because the specific option, which was part of the CONFIG_PREEMPT_VOLUNTARY kernel config option and that allowed for setting non-threaded IRQs, simply doesn't exist anymore. So basically this variable doesn't do anything.

The RTIRQ_HIGH_LIST variable contains a list of space separated service names that get priorities in the 99-91 range, so above the value as set in the RTIRQ_PRIO_HIGH variable. Use this variable only for services of which you want to be 100% sure they don't get interrupted by anything else. You will mostly want to put timers in there like rtc or the ALSA high resolution timer (snd-hrtimer).

When to use rtirq

A common use case for rtirq is when a FireWire controller shares its IRQ with one or even more devices. With rtirq you can then give a higher priority to the FireWire bottom-halve which should lower the chance of running into xruns drastically or even enable the device to initialize properly. The grep firewire /proc/interrupts command outputs the IRQ that is in use by your FireWire controller (when using the current JuJu stack):

$ grep firewire /proc/interrupts 
16:      88504        215   IO-APIC-fasteoi   uhci_hcd:usb3, mmc0, firewire_ohci, jmb38x_ms:slot0, eth1, nvidia

As you can see IRQ 16 is really crowded in this case and a FireWire device connected to this controller simply won't initialize. By adding the firewire_ohci module to the list of separated values of the RTIRQ_NAME_LIST variable you can assign a higher priority to the softirq or bottom-halve of the FireWire controller:

RTIRQ_NAME_LIST="rtc firewire usb"

If RTIRQ_PRIO_HIGH is set to 90 and RTIRQ_PRIO_DECR to 5 then the real-time clock (rtc) will get prio 90 and the FireWire controller prio 85. All other bottom-halves of IRQ 16 will get a default prio of 50 so the FireWire controller will always take precedence. In this particular case (a HP DV7-1040) using rtirq this way enables the FireWire device (a Focusrite Saffire Pro 10 IO) to initialize properly and run in a stable manner at lower latencies (128 frames/48Khz samplerate * 3 periods = 8ms). And given the fact that the controller (an onboard JMicron chipset) sits on a very crowded IRQ shared by a NIC, a cardreader, an USB port and a GPU rtirq definitely shows off its strength here.

PCI bus latency

Conventional PCI has a shared bus topology and also suffers from only being able to address a limited number of IRQ's. So besides PCI devices sharing interrupts they could also get into each other's way because of

lengthy transactions by some devices, due to slow acknowledgments, long data bursts, or some combination, which could cause buffer underrun or overrun in other devices”4). It's successor, PCI-X, suffers less from PCI bus latency because it uses higher bandwidths, but it's still using a shared bus topology unlike the latest PCI standard, PCIe. PCIe is “based on point-to-point topology, with separate serial links connecting every device to the host”5). > >So if you're using conventional PCI sound cards, you might want to consider maximizing the PCI latency timer of the sound card and raising the latency timer of all other PCI peripherals a bit (default is 64). This may sound counter-intuitive but “the PCI latency timer determines how long a device may own the PCI bus for … Therefore, if your system allowed the video interface to hog the PCI bus for a relatively long time, but limited the audio interface to only short periods of bus ownership, then audio performance will suffer. > >The script … resets *all* devices to an acceptable middle-of-the-road timer setting, then specifically allows the audio interfaces to own the bus for even longer.6)
#!/bin/sh

case $1 in
     start)
           # "open up" the PCI bus by allowing fairly long bursts for all devices, increasing performance
           setpci -v -s "*:*.*" latency_timer=b0

           # maximise latency for RME Hammerfall, thus allowing
           # more data per PCI transfer and minimizing xruns
           setpci -v -s 01:04.0 latency_timer=ff
           # ditto for the onboard AC97 audio interface
           setpci -v -s 00:07.5 latency_timer=ff
esac

exit 0

Source of the script: http://lists.linuxaudio.org/pipermail/linux-audio-user/2007-April/044371.html

Question remains in how far adjusting the PCI latency timer affects performance for PCI-X devices. PCIe doesn't suffer from PCI bus latency because of its different topology. Besides, PCIe devices have their PCI latency timer set to 0 and the setpci command simply doesn't change anything, setting latency_timer=ff on PCIe devices still yields a setting of 0.

Screensavers

When doing audio work screensavers tend to get in the way. So you might want to consider uninstalling them. You will find then that X is still trying to blank your screen after a while. To completely disable this you could add the following lines to your .profile file:

xset -dpms
xset s off
setterm -powersave off -blank 0

This should completely disable any unwanted screensavers or screen blanking.

PulseAudio

Basically all modern distros come with PulseAudio installed. PulseAudio (PA) is a sound server (like JACK) that takes care of mixing the output of all applications that can output sound. Since PulseAudio is designed to work in as much situations and with as much hardware as possible it is not the best choice for a real-time, low-latency Linux audio set-up. In other words PulseAudio is more suited for consumer audio and JACK is more suited for pro audio.

There are distros that offer scripts or configurations to allow using both sound servers at once. But when setting up a Linux audio workstation you could ask yourself if you really need two sound daemons when the software you're going to use is probably all JACK aware. Also, think about the extra overhead and possible issues you could run into when using two sound servers. Hence it's not uncommon that people disable or uninstall PulseAudio on their Linux audio machines. On a Linux audio workstation that is used for nothing else you could uninstall PulseAudio, on a workstation that is also used for other tasks it might be a better option to disable PulseAudio for the specific session in which you want to do your audio and music production. Since PulseAudio is the default sound server on a lot of distros it is not part of the startup applications and it is also not called through init systems. So it will always run or start to run as soon as an application wants to output sound. Add to this that on most distros PA is set to respawn automatically after being killed or when it crashes and you'll understand that PA doesn't allow to be disabled easily.

There is a relatively easy way to disable PulseAudio though for a specific user session. First, modify or create the file ~/.pulse/client.conf and make sure it contains the following line:

autospawn = no

Now you can kill the pulseaudio daemon:

killall -9 pulseaudio

Or with:

pulseaudio -k

And it shouldn't come up again, providing a PA free user session.

Using FireWire

When using FireWire audio interfaces it is generally advised to use a real-time kernel. Of course this is not necessary, there are also users reporting success with generic kernels, but when you have trouble getting your system stable regarding xruns one of the first things besides installing rtirq you might want to try is to install or compile a real-time kernel.

Legacy FireWire stack

Make sure the raw1394 module is loaded. You can have this module loaded automatically by adding it to your /etc/modules file. Just add the following line to this file:

raw1394

Loading this module will create the /dev/raw1394 node. Normal users do not have permissions to use this node but with an udev rule in /etc/udev/rules.d/ you can control these permissions. Most recent distributions also install the corresponding udev rules when installing FFADO packages but if your distro doesn't do so you can create a new file 50-raw-firewire.rules with the following line to allow the audio group to use this node:

KERNEL=="raw1394", GROUP="audio"

You will also need the FFADO drivers (most distros have packages for these drivers) and set the driver for JACK to 'firewire' in either QjackCtl or on the commandline with the -dfirewire argument.

JuJu FireWire stack

As of kernel 2.6.37 the legacy FireWire stack has been removed from the kernel sources in favor of the JuJu stack. If you have FFADO installed it should take care of properly setting up your device(s). Instead of creating a /dev/raw1394 node, the JuJu stack creates nodes per device in /dev/, first device is always /dev/fw0. Take note that the first device is always your controller, not your soundcard.

To use JACK with your FireWire device(s), set the driver to 'firewire' in QjackCtl or use the -dfirewire option on the commandline.

Troubleshooting

After setting up, configuring and tweaking your system it is still possible that it doesn't perform the way you want it to. To help analyzing why the overal performance is not as expected there are some helpful tools that might reveal the source of the bottlenecks on your system.

top

top is probably the most used tool to display system information like CPU or memory usage. It is installed by default on most distros.

Terminal running top

With the help of top you can find out which process is using up all your CPU or memory. By pressing h you get a help screen. To sort on memory usage instead of CPU usage you can press F + n + Enter.

htop

htop is the somewhat more sophisticated big brother of top.

Terminal running htop

latencytop

latencytop is a very useful tool to find out what process is causing system latency to happen.

latencytop

cyclictest

Just as latencytop measures system latencies, cyclictest measures kernel latencies. cyclictest has quite some options but one of the most used ways to run the command is:

#  cyclictest -t1 -p 80 -n -i 10000 -l 10000 -m

One single thread (-t1), priority of 80 (-p 80), use clock_nanosleep (-n), use 10000 us base interval of thread (-i 1000), use 1000 loops and then exit (-l 10000), lock current and future memory allocations (-m). More on clock_nanosleep can be found in its manpage (man clock_nanosleep).

Other examples of the usage of cyclictest:

TODO

  • The snd-hrtimer module
  • More on MIDI timing
  • Harddrive tuning (does this even apply for SATA drives??)
  • Troubleshooting tools (ffado-diag, alsa latency tool, alsa midi latency tool, jack_delay, tuna)
wiki/system_configuration.1496052668.txt.gz · Last modified: 2017/05/29 12:11 by autostatic