10 Kernel control groups #
Kernel Control Groups (“cgroups”) are a kernel feature for assigning and limiting hardware and system resources for processes. Processes can also be organized in a hierarchical tree structure.
10.1 Overview #
Every process is assigned exactly one administrative cgroup. cgroups are ordered in a hierarchical tree structure. You can set resource limitations, such as CPU, memory, disk I/O, or network bandwidth usage, for single processes or for whole branches of the hierarchy tree.
On SUSE Linux Enterprise Server, systemd
uses cgroups to organize all processes in
groups, which systemd
calls slices. systemd
also provides an
interface for setting cgroup properties.
The command systemd-cgls
displays the hierarchy tree.
The kernel cgroup API comes in two variants, v1 and v2. Additionally, there can be multiple cgroup hierarchies exposing different APIs. From the numerous possible combinations, there are two practical choices:
hybrid: v2 hierarchy without controllers, and the controllers are on v1 hierarchies
unified: v2 hierarchy with controllers
The current default mode is hybrid. This provides backwards compatibility for applications that need it. The following features are available only with the unified v2 hierarchy:
memory controller: reclaim protection (aka memory.low), memory.high, PSI (pressure stall information)
io controller: writeback control, new control policies
controller delegation to non-privileged users (rootless containers)
freezer support in systemd
simpler handling of the single hierararchy
You may set only one mode.
To enable the unified control group hierarchy, append
systemd.unified_cgroup_hierarchy=1
as a kernel command
line parameter to the GRUB 2 boot loader. (Refer to
Chapter 18, The boot loader GRUB 2 for more details about configuring GRUB 2.)
10.2 Resource accounting #
Organizing processes into different cgroups can be used to obtain per-cgroup resource consumption data.
The accounting has relatively small but non-zero overhead, whose impact depends on the workload. Activating accounting for one unit will also implicitly activate it for all units in the same slice, and for all its parent slices, and the units contained in them.
The accounting can be set on a per-unit basis with directives such as
MemoryAccounting=
or globally for all units in
/etc/systemd/system.conf
with the directive
DefaultMemoryAccounting=
. Refer to man
systemd.resource-control
for the exhaustive list of possible
directives.
10.3 Setting resource limits #
Be aware that resource consumption implicitly depends on the environment where your workload executes (for example, size of data structures in libraries/kernel, forking behavior of utilities, computational efficiency). Hence it is recommended to (re)calibrate your limits should the environment change.
Limitations to cgroups can be set with the systemctl
set-property
command. The syntax is:
#
systemctl set-property [--runtime] NAME PROPERTY1=VALUE [PROPERTY2=VALUE]
The configured value is applied immediately. Optionally, use the
--runtime
option, so that the new values do not persist
after reboot.
Replace NAME with a systemd
service, scope, or
slice name.
For a complete list of properties and more details, see man
systemd.resource-control
.
10.4 Preventing fork bombs with TasksMax
#
systemd
supports configuring task count limits both for each individual
leaf unit, or aggregated on slices. Upstream systemd
ships with defaults
that limit the number of tasks in each unit (15% of the kernel global limit,
run /usr/sbin/sysctl kernel.pid_max
to see the total
limit). Each user's slice is limited to 33% of the kernel limit. However,
this is different for SUSE Linux Enterprise Server.
10.4.1 Finding the current default TasksMax
values #
It became apparent, in practice, that there is not a single default that
applies to all use cases. SUSE Linux Enterprise Server ships with two custom
configurations that override the upstream defaults for system units and for
user slices, and sets them both to infinity
.
/usr/lib/systemd/system.conf.d/__25-defaults-SLE.conf
contains these lines:
[Manager] DefaultTasksMax=infinity
/usr/lib/systemd/system/user-.slice.d/25-defaults-SLE.conf
contains these lines:
[Slice] TasksMax=infinity
Use systemctl
to verify the DefaultTasksMax value:
>
systemctl show --property DefaultTasksMax
DefaultTasksMax=infinity
infinity
means having no limit. It is not a requirement
to change the default, but setting some limits may help to prevent system
crashes from runaway processes.
10.4.2 Overriding the DefaultTasksMax
value #
Change the global DefaultTasksMax
value by creating a
new override file,
/etc/systemd/system.conf.d/90-system-tasksmax.conf
,
and write the following lines to set a new default limit of 256 tasks per
system unit:
[Manager] DefaultTasksMax=256
Load the new setting, then verify that it changed:
>
sudo
systemctl daemon-reload
>
systemctl show --property DefaultTasksMax
DefaultTasksMax=256
Adjust this default value to suit your needs. You can set different limits on individual services as needed. This example is for MariaDB. First check the current active value:
>
systemctl status mariadb.service
● mariadb.service - MariaDB database server Loaded: loaded (/usr/lib/systemd/system/mariadb.service; disabled; vendor preset> Active: active (running) since Tue 2020-05-26 14:15:03 PDT; 27min ago Docs: man:mysqld(8) https://mariadb.com/kb/en/library/systemd/ Main PID: 11845 (mysqld) Status: "Taking your SQL requests now..." Tasks: 30 (limit: 256) CGroup: /system.slice/mariadb.service └─11845 /usr/sbin/mysqld --defaults-file=/etc/my.cnf --user=mysql
The Tasks line shows that MariaDB currently has 30 tasks running, and has an upper limit of the default 256, which is inadequate for a database. The following example demonstrates how to raise MariaDB's limit to 8192.
>
sudo
systemctl set-property mariadb.service TasksMax=8192
>
systemctl status mariadb.service
● mariadb.service - MariaDB database server Loaded: loaded (/usr/lib/systemd/system/mariadb.service; disabled; vendor preset: disab> Drop-In: /etc/systemd/system/mariadb.service.d └─50-TasksMax.conf Active: active (running) since Tue 2020-06-02 17:57:48 PDT; 7min ago Docs: man:mysqld(8) https://mariadb.com/kb/en/library/systemd/ Process: 3446 ExecStartPre=/usr/lib/mysql/mysql-systemd-helper upgrade (code=exited, sta> Process: 3440 ExecStartPre=/usr/lib/mysql/mysql-systemd-helper install (code=exited, sta> Main PID: 3452 (mysqld) Status: "Taking your SQL requests now..." Tasks: 30 (limit: 8192) CGroup: /system.slice/mariadb.service └─3452 /usr/sbin/mysqld --defaults-file=/etc/my.cnf --user=mysql
systemctl set-property
applies the new limit and creates
a drop-in file for persistence,
/etc/systemd/system/mariadb.service.d/50-TasksMax.conf
,
that contains only the changes you want to apply to the existing unit file.
The value does not have to be 8192, but should be whatever limit is
appropriate for your workloads.
10.4.3 Default TasksMax
limit on users #
The default limit on users should be fairly high, because user sessions
need more resources. Set your own default for any user by creating a new
file, for example
/etc/systemd/system/user-.slice.d/40-user-taskmask.conf
.
The following example sets a default of 16284:
[Slice] TasksMax=16284
See Section 19.5.3, “Creating drop-in files manually” to learn what numeric prefixes are expected for drop-in files.
Then reload systemd to load the new value, and verify the change:
>
sudo
systemctl daemon-reload
>
systemctl show --property TasksMax user-1000.slice
TasksMax=16284
How do you know what values to use? This varies according to your
workloads, system resources, and other resource configurations. When your
TasksMax
value is too low, you will see error messages
such as Failed to fork (Resources temporarily
unavailable), Can't create thread to handle new
connection, and Error: Function call 'fork' failed
with error code 11, 'Resource temporarily unavailable'.
For more information on configuring system resources in systemd, see
systemd.resource-control (5)
.
10.5 I/O control with cgroups #
This section introduces using the Linux kernel's block I/O controller to prioritize or throttle I/O operations. This leverages the means provided by systemd to configure cgroups, and discusses probable pitfalls when dealing with proportional I/O control.
10.5.1 Prerequisites #
The following subsections describe steps that you must take in advance when you design and configure your system, since those aspects cannot be changed during runtime.
10.5.1.1 File system #
You should use a cgroup-writeback-aware file system (otherwise writeback charging is not possible). The recommended SUSE Linux Enterprise Server file systems added support in the following upstream releases:
Btrfs (v4.3)
Ext4 (v4.3)
XFS (v5.3)
As of SUSE Linux Enterprise Server 15 SP3, any of the named file systems can be used.
10.5.1.2 Unified cgroup hierarchy #
To properly account writeback I/O, it is necessary to have equal I/O
and memory controller cgroup hierarchies, and to use the cgroup v2 I/O
controller. Together, this means that one has to use the
unified cgroup hierarchy. This has to be requested
explicitly in SUSE Linux Enterprise Server by passing a kernel command line option,
systemd.unified_cgroup_hierarchy=1
.
10.5.1.3 Block I/O scheduler #
The throttling policy is implemented higher in the stack, therefore it does not require any additional adjustments. The proportional I/O control policies have two different implementations: the BFQ controller, and the cost-based model. We describe the BFQ controller here. In order to exert its proportional implementation for a particular device, we must make sure that BFQ is the chosen scheduler. Check the current scheduler:
>
cat /sys/class/block/sda/queue/scheduler
mq-deadline kyber bfq [none]
Switch the scheduler to BFQ:
#
echo bfq > /sys/class/block/sda/queue/scheduler
You must specify the disk device (not a partition). The optimal way to set this attribute is a udev rule specific to the device (note that SUSE Linux Enterprise Server ships udev rules that already enable BFQ for rotational disk drives).
10.5.1.4 Cgroup hierarchy layout #
Normally, all tasks reside in the root cgroup and they compete against each other. When the tasks are distributed into the cgroup tree the competition occurs between sibling cgroups only. This applies to the proportional I/O control; the throttling hierarchically aggregates throughput of all descendants (see the following diagram).
r `- a IOWeight=100 `- [c] IOWeight=300 `- d IOWeight=100 `- [b] IOWeight=200
I/O is originating only from cgroups c and b. Even though c has a higher weight, it will be treated with lower priority because it is level-competing with b.
10.5.2 Configuring control quantities #
You can apply the values to (long running) services permanently.
>
sudo
systemctl set-property fast.service IOWeight=400
>
sudo
systemctl set-property slow.service IOWeight=50
>
sudo
systemctl set-property throttled.service IOReadBandwidthMax="/dev/sda 1M"
Alternatively, you can apply I/O control to individual commands, for example:
>
sudo
systemd-run --scope -p IOWeight=400 high_prioritized_command
>
sudo
systemd-run --scope -p IOWeight=50 low_prioritized_command
>
sudo
systemd-run --scope -p IOReadBandwidthMax="/dev/sda 1M" dd if=/dev/sda of=/dev/null bs=1M count=10
10.5.3 I/O control behavior and setting expectations #
The following list items describe I/O control behavior, and what you should expect under various conditions.
I/O control works best for direct I/O operations (bypassing page cache), the situations where the actual I/O is decoupled from the caller (typically writeback via page cache) may manifest variously. For example, delayed I/O control or even no observed I/O control (consider little bursts or competing workloads that happen to never "meet", submitting I/O at the same time, and saturating the bandwidth). For these reasons, the resulting ratio of I/O throughputs does not strictly follow the ratio of configured weights.
systemd performs scaling of configured weights (to adjust for narrower BFQ weight range), hence the resulting throughput ratios also differ.
The writeback activity depends on the amount of dirty pages, besides the global sysctl knobs (
vm.dirty_background_ratio
andvm.dirty_ratio
)). Memory limits of individual cgroups come into play when the dirty limits are distributed among cgroups, and this in turn may affect I/O intensity of affected cgroups.Not all storages are equal. The I/O control happens at the I/O scheduler layer, which has ramifications for setups with devices stacked on these that do no actual scheduling. Consider device mapper logical volumes spanning multiple physical devices, MD RAID, or even Btrfs RAID. I/O control over such setups may be challenging.
There is no separate setting for proportional I/O control of reads and writes.
Proportional I/O control is only one of the policies that can interact with each other (but responsible resource design perhaps avoids that).
The I/O device bandwidth is not the only shared resource on the I/O path. Global file system structures are involved, which is relevant when I/O control is meant to guarantee certain bandwidth; it will not, and it may even lead to priority inversion (prioritized cgroup waiting for a transaction of slower cgroup).
So far, we have been discussing only explicit I/O of file system data, but swap-in and swap-out can also be controlled. Although if such a need arises, it usually points out to improperly provisioned memory (or memory limits).
10.6 More information #
Kernel documentation (package
kernel-source
): files in/usr/src/linux/Documentation/admin-guide/cgroup-v1
and file/usr/src/linux/Documentation/admin-guide/cgroup-v2.rst
.man systemd.resource-control
https://lwn.net/Articles/604609/—Brown, Neil: Control Groups Series (2014, 7 parts).
https://lwn.net/Articles/243795/—Corbet, Jonathan: Controlling memory use in containers (2007).
https://lwn.net/Articles/236038/—Corbet, Jonathan: Process containers (2007).