libvirt/docs/cgroups.html.in

421 lines
14 KiB
XML

<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE html>
<html xmlns="http://www.w3.org/1999/xhtml">
<body>
<h1>Control Groups Resource Management</h1>
<ul id="toc"></ul>
<p>
The QEMU and LXC drivers make use of the Linux "Control Groups" facility
for applying resource management to their virtual machines and containers.
</p>
<h2><a id="requiredControllers">Required controllers</a></h2>
<p>
The control groups filesystem supports multiple "controllers". By default
the init system (such as systemd) should mount all controllers compiled
into the kernel at <code>/sys/fs/cgroup/$CONTROLLER-NAME</code>. Libvirt
will never attempt to mount any controllers itself, merely detect where
they are mounted.
</p>
<p>
The QEMU driver is capable of using the <code>cpuset</code>,
<code>cpu</code>, <code>cpuacct</code>, <code>memory</code>,
<code>blkio</code> and <code>devices</code> controllers.
None of them are compulsory. If any controller is not mounted,
the resource management APIs which use it will cease to operate.
It is possible to explicitly turn off use of a controller,
even when mounted, via the <code>/etc/libvirt/qemu.conf</code>
configuration file.
</p>
<p>
The LXC driver is capable of using the <code>cpuset</code>,
<code>cpu</code>, <code>cpuacct</code>, <code>freezer</code>,
<code>memory</code>, <code>blkio</code> and <code>devices</code>
controllers. The <code>cpuacct</code>, <code>devices</code>
and <code>memory</code> controllers are compulsory. Without
them mounted, no containers can be started. If any of the
other controllers are not mounted, the resource management APIs
which use them will cease to operate.
</p>
<h2><a id="currentLayout">Current cgroups layout</a></h2>
<p>
As of libvirt 1.0.5 or later, the cgroups layout created by libvirt has been
simplified, in order to facilitate the setup of resource control policies by
administrators / management applications. The new layout is based on the concepts
of "partitions" and "consumers". A "consumer" is a cgroup which holds the
processes for a single virtual machine or container. A "partition" is a cgroup
which does not contain any processes, but can have resource controls applied.
A "partition" will have zero or more child directories which may be either
"consumer" or "partition".
</p>
<p>
As of libvirt 1.1.1 or later, the cgroups layout will have some slight
differences when running on a host with systemd 205 or later. The overall
tree structure is the same, but there are some differences in the naming
conventions for the cgroup directories. Thus the following docs split
in two, one describing systemd hosts and the other non-systemd hosts.
</p>
<h3><a id="currentLayoutSystemd">Systemd cgroups integration</a></h3>
<p>
On hosts which use systemd, each consumer maps to a systemd scope unit,
while partitions map to a system slice unit.
</p>
<h4><a id="systemdScope">Systemd scope naming</a></h4>
<p>
The systemd convention is for the scope name of virtual machines / containers
to be of the general format <code>machine-$NAME.scope</code>. Libvirt forms the
<code>$NAME</code> part of this by concatenating the driver type with the id
and truncated name of the guest, and then escaping any systemd reserved
characters.
So for a guest <code>demo</code> running under the <code>lxc</code> driver,
we get a <code>$NAME</code> of <code>lxc-12345-demo</code> which when escaped
is <code>lxc\x2d12345\x2ddemo</code>. So the complete scope name is
<code>machine-lxc\x2d12345\x2ddemo.scope</code>.
The scope names map directly to the cgroup directory names.
</p>
<h4><a id="systemdSlice">Systemd slice naming</a></h4>
<p>
The systemd convention for slice naming is that a slice should include the
name of all of its parents prepended on its own name. So for a libvirt
partition <code>/machine/engineering/testing</code>, the slice name will
be <code>machine-engineering-testing.slice</code>. Again the slice names
map directly to the cgroup directory names. Systemd creates three top level
slices by default, <code>system.slice</code> <code>user.slice</code> and
<code>machine.slice</code>. All virtual machines or containers created
by libvirt will be associated with <code>machine.slice</code> by default.
</p>
<h4><a id="systemdLayout">Systemd cgroup layout</a></h4>
<p>
Given this, a possible systemd cgroups layout involving 3 qemu guests,
3 lxc containers and 3 custom child slices, would be:
</p>
<pre>
$ROOT
|
+- system.slice
| |
| +- libvirtd.service
|
+- machine.slice
|
+- machine-qemu\x2d1\x2dvm1.scope
| |
| +- emulator
| +- vcpu0
| +- vcpu1
|
+- machine-qemu\x2d2\x2dvm2.scope
| |
| +- emulator
| +- vcpu0
| +- vcpu1
|
+- machine-qemu\x2d3\x2dvm3.scope
| |
| +- emulator
| +- vcpu0
| +- vcpu1
|
+- machine-engineering.slice
| |
| +- machine-engineering-testing.slice
| | |
| | +- machine-lxc\x2d11111\x2dcontainer1.scope
| |
| +- machine-engineering-production.slice
| |
| +- machine-lxc\x2d22222\x2dcontainer2.scope
|
+- machine-marketing.slice
|
+- machine-lxc\x2d33333\x2dcontainer3.scope
</pre>
<h3><a id="currentLayoutGeneric">Non-systemd cgroups layout</a></h3>
<p>
On hosts which do not use systemd, each consumer has a corresponding cgroup
named <code>$VMNAME.libvirt-{qemu,lxc}</code>. Each consumer is associated
with exactly one partition, which also have a corresponding cgroup usually
named <code>$PARTNAME.partition</code>. The exceptions to this naming rule
are the three top level default partitions, named <code>/system</code> (for
system services), <code>/user</code> (for user login sessions) and
<code>/machine</code> (for virtual machines and containers). By default
every consumer will of course be associated with the <code>/machine</code>
partition.
</p>
<p>
Given this, a possible systemd cgroups layout involving 3 qemu guests,
3 lxc containers and 2 custom child slices, would be:
</p>
<pre>
$ROOT
|
+- system
| |
| +- libvirtd.service
|
+- machine
|
+- qemu-1-vm1.libvirt-qemu
| |
| +- emulator
| +- vcpu0
| +- vcpu1
|
+- qeme-2-vm2.libvirt-qemu
| |
| +- emulator
| +- vcpu0
| +- vcpu1
|
+- qemu-3-vm3.libvirt-qemu
| |
| +- emulator
| +- vcpu0
| +- vcpu1
|
+- engineering.partition
| |
| +- testing.partition
| | |
| | +- lxc-11111-container1.libvirt-lxc
| |
| +- production.partition
| |
| +- lxc-22222-container2.libvirt-lxc
|
+- marketing.partition
|
+- lxc-33333-container3.libvirt-lxc
</pre>
<h2><a id="customPartiton">Using custom partitions</a></h2>
<p>
If there is a need to apply resource constraints to groups of
virtual machines or containers, then the single default
partition <code>/machine</code> may not be sufficiently
flexible. The administrator may wish to sub-divide the
default partition, for example into "testing" and "production"
partitions, and then assign each guest to a specific
sub-partition. This is achieved via a small element addition
to the guest domain XML config, just below the main <code>domain</code>
element
</p>
<pre>
...
&lt;resource&gt;
&lt;partition&gt;/machine/production&lt;/partition&gt;
&lt;/resource&gt;
...
</pre>
<p>
Note that the partition names in the guest XML are using a
generic naming format, not the low level naming convention
required by the underlying host OS. That is, you should not include
any of the <code>.partition</code> or <code>.slice</code>
suffixes in the XML config. Given a partition name
<code>/machine/production</code>, libvirt will automatically
apply the platform specific translation required to get
<code>/machine/production.partition</code> (non-systemd)
or <code>/machine.slice/machine-production.slice</code>
(systemd) as the underlying cgroup name
</p>
<p>
Libvirt will not auto-create the cgroups directory to back
this partition. In the future, libvirt / virsh will provide
APIs / commands to create custom partitions, but currently
this is left as an exercise for the administrator.
</p>
<p>
<strong>Note:</strong> the ability to place guests in custom
partitions is only available with libvirt &gt;= 1.0.5, using
the new cgroup layout. The legacy cgroups layout described
later in this document did not support customization per guest.
</p>
<h3><a id="createSystemd">Creating custom partitions (systemd)</a></h3>
<p>
Given the XML config above, the admin on a systemd based host would
need to create a unit file <code>/etc/systemd/system/machine-production.slice</code>
</p>
<pre>
# cat &gt; /etc/systemd/system/machine-testing.slice &lt;&lt;EOF
[Unit]
Description=VM testing slice
Before=slices.target
Wants=machine.slice
EOF
# systemctl start machine-testing.slice
</pre>
<h3><a id="createNonSystemd">Creating custom partitions (non-systemd)</a></h3>
<p>
Given the XML config above, the admin on a non-systemd based host
would need to create a cgroup named '/machine/production.partition'
</p>
<pre>
# cd /sys/fs/cgroup
# for i in blkio cpu,cpuacct cpuset devices freezer memory net_cls perf_event
do
mkdir $i/machine/production.partition
done
# for i in cpuset.cpus cpuset.mems
do
cat cpuset/machine/$i > cpuset/machine/production.partition/$i
done
</pre>
<h2><a id="resourceAPIs">Resource management APIs/commands</a></h2>
<p>
Since libvirt aims to provide an API which is portable across
hypervisors, the concept of cgroups is not exposed directly
in the API or XML configuration. It is considered to be an
internal implementation detail. Instead libvirt provides a
set of APIs for applying resource controls, which are then
mapped to corresponding cgroup tunables
</p>
<h3>Scheduler tuning</h3>
<p>
Parameters from the "cpu" controller are exposed via the
<code>schedinfo</code> command in virsh.
</p>
<pre>
# virsh schedinfo demo
Scheduler : posix
cpu_shares : 1024
vcpu_period : 100000
vcpu_quota : -1
emulator_period: 100000
emulator_quota : -1</pre>
<h3>Block I/O tuning</h3>
<p>
Parameters from the "blkio" controller are exposed via the
<code>bkliotune</code> command in virsh.
</p>
<pre>
# virsh blkiotune demo
weight : 500
device_weight : </pre>
<h3>Memory tuning</h3>
<p>
Parameters from the "memory" controller are exposed via the
<code>memtune</code> command in virsh.
</p>
<pre>
# virsh memtune demo
hard_limit : 580192
soft_limit : unlimited
swap_hard_limit: unlimited
</pre>
<h3>Network tuning</h3>
<p>
The <code>net_cls</code> is not currently used. Instead traffic
filter policies are set directly against individual virtual
network interfaces.
</p>
<h2><a id="legacyLayout">Legacy cgroups layout</a></h2>
<p>
Prior to libvirt 1.0.5, the cgroups layout created by libvirt was different
from that described above, and did not allow for administrator customization.
Libvirt used a fixed, 3-level hierarchy <code>libvirt/{qemu,lxc}/$VMNAME</code>
which was rooted at the point in the hierarchy where libvirtd itself was
located. So if libvirtd was placed at <code>/system/libvirtd.service</code>
by systemd, the groups for each virtual machine / container would be located
at <code>/system/libvirtd.service/libvirt/{qemu,lxc}/$VMNAME</code>. In addition
to this, the QEMU drivers further child groups for each vCPU thread and the
emulator thread(s). This leads to a hierarchy that looked like
</p>
<pre>
$ROOT
|
+- system
|
+- libvirtd.service
|
+- libvirt
|
+- qemu
| |
| +- vm1
| | |
| | +- emulator
| | +- vcpu0
| | +- vcpu1
| |
| +- vm2
| | |
| | +- emulator
| | +- vcpu0
| | +- vcpu1
| |
| +- vm3
| |
| +- emulator
| +- vcpu0
| +- vcpu1
|
+- lxc
|
+- container1
|
+- container2
|
+- container3
</pre>
<p>
Although current releases are much improved, historically the use of deep
hierarchies has had a significant negative impact on the kernel scalability.
The legacy libvirt cgroups layout highlighted these problems, to the detriment
of the performance of virtual machines and containers.
</p>
</body>
</html>