nixpkgs/nixos/doc/manual/administration/control-groups.xml

<chapter xmlns="http://docbook.org/ns/docbook"
        xmlns:xlink="http://www.w3.org/1999/xlink"
        xmlns:xi="http://www.w3.org/2001/XInclude"
        version="5.0"
        xml:id="sec-cgroups">
 <title>Control Groups</title>
 <para>
  To keep track of the processes in a running system, systemd uses
  <emphasis>control groups</emphasis> (cgroups). A control group is a set of
  processes used to allocate resources such as CPU, memory or I/O bandwidth.
  There can be multiple control group hierarchies, allowing each kind of
  resource to be managed independently.
 </para>
 <para>
  The command <command>systemd-cgls</command> lists all control groups in the
  <literal>systemd</literal> hierarchy, which is what systemd uses to keep
  track of the processes belonging to each service or user session:
<screen>
<prompt>$ </prompt>systemd-cgls
├─user
│ └─eelco
│   └─c1
│     ├─ 2567 -:0
│     ├─ 2682 kdeinit4: kdeinit4 Running...
│     ├─ <replaceable>...</replaceable>
│     └─10851 sh -c less -R
└─system
  ├─httpd.service
  │ ├─2444 httpd -f /nix/store/3pyacby5cpr55a03qwbnndizpciwq161-httpd.conf -DNO_DETACH
  │ └─<replaceable>...</replaceable>
  ├─dhcpcd.service
  │ └─2376 dhcpcd --config /nix/store/f8dif8dsi2yaa70n03xir8r653776ka6-dhcpcd.conf
  └─ <replaceable>...</replaceable>
</screen>
  Similarly, <command>systemd-cgls cpu</command> shows the cgroups in the CPU
  hierarchy, which allows per-cgroup CPU scheduling priorities. By default,
  every systemd service gets its own CPU cgroup, while all user sessions are in
  the top-level CPU cgroup. This ensures, for instance, that a thousand
  run-away processes in the <literal>httpd.service</literal> cgroup cannot
  starve the CPU for one process in the <literal>postgresql.service</literal>
  cgroup. (By contrast, it they were in the same cgroup, then the PostgreSQL
  process would get 1/1001 of the cgroup’s CPU time.) You can limit a
  service’s CPU share in <filename>configuration.nix</filename>:
<programlisting>
<link linkend="opt-systemd.services._name_.serviceConfig">systemd.services.httpd.serviceConfig</link>.CPUShares = 512;
</programlisting>
  By default, every cgroup has 1024 CPU shares, so this will halve the CPU
  allocation of the <literal>httpd.service</literal> cgroup.
 </para>
 <para>
  There also is a <literal>memory</literal> hierarchy that controls memory
  allocation limits; by default, all processes are in the top-level cgroup, so
  any service or session can exhaust all available memory. Per-cgroup memory
  limits can be specified in <filename>configuration.nix</filename>; for
  instance, to limit <literal>httpd.service</literal> to 512 MiB of RAM
  (excluding swap):
<programlisting>
<link linkend="opt-systemd.services._name_.serviceConfig">systemd.services.httpd.serviceConfig</link>.MemoryLimit = "512M";
</programlisting>
 </para>
 <para>
  The command <command>systemd-cgtop</command> shows a continuously updated
  list of all cgroups with their CPU and memory usage.
 </para>
</chapter>