162 lines
5.6 KiB
ReStructuredText
162 lines
5.6 KiB
ReStructuredText
.. _config-ovs-dpdk:
|
|
|
|
===============================
|
|
Open vSwitch with DPDK datapath
|
|
===============================
|
|
|
|
This page serves as a guide for how to use the OVS with DPDK datapath
|
|
functionality available in the Networking service as of the Mitaka release.
|
|
|
|
The basics
|
|
~~~~~~~~~~
|
|
|
|
Open vSwitch (OVS) provides support for a Data Plane Development Kit (DPDK)
|
|
datapath since OVS 2.2, and a DPDK-backed ``vhost-user`` virtual interface
|
|
since OVS 2.4. The DPDK datapath provides lower latency and higher performance
|
|
than the standard kernel OVS datapath, while DPDK-backed ``vhost-user``
|
|
interfaces can connect guests to this datapath. For more information on DPDK,
|
|
refer to the `DPDK <http://dpdk.org/>`__ website.
|
|
|
|
OVS with DPDK, or OVS-DPDK, can be used to provide high-performance networking
|
|
between instances on OpenStack compute nodes.
|
|
|
|
Prerequisites
|
|
-------------
|
|
|
|
Using DPDK in OVS requires the following minimum software versions:
|
|
|
|
* OVS 2.4
|
|
* DPDK 2.0
|
|
* QEMU 2.1.0
|
|
* libvirt 1.2.13
|
|
|
|
Support of ``vhost-user`` multiqueue that enables use of multiqueue with
|
|
``virtio-net`` and ``igb_uio`` is available if the following newer
|
|
versions are used:
|
|
|
|
* OVS 2.5
|
|
* DPDK 2.2
|
|
* QEMU 2.5
|
|
* libvirt 1.2.17
|
|
|
|
In both cases, install and configure Open vSwitch with DPDK support for each
|
|
node. For more information, see the
|
|
`OVS-DPDK <https://github.com/openvswitch/ovs/blob/master/Documentation/intro/install/dpdk.rst>`__
|
|
installation guide (select an appropriate OVS version in the
|
|
:guilabel:`Branch` drop-down menu).
|
|
|
|
:doc:`/contributor/internals/ovs_vhostuser`
|
|
for configuration of neutron OVS agent.
|
|
|
|
In case you wish to configure multiqueue, see the
|
|
`OVS configuration chapter on vhost-user
|
|
<http://wiki.qemu.org/Documentation/vhost-user-ovs-dpdk#Enabling_multi-queue>`__
|
|
in QEMU documentation.
|
|
|
|
The technical background of multiqueue is explained in the corresponding
|
|
`blueprint <https://specs.openstack.org/openstack/nova-specs/specs/liberty/implemented/libvirt-virtiomq.html>`__.
|
|
|
|
Additionally, OpenStack supports ``vhost-user`` reconnect feature starting
|
|
from the Ocata release, as implementation of fix for
|
|
`bug 1604924 <https://bugs.launchpad.net/neutron/+bug/1604924>`__.
|
|
Starting from OpenStack Ocata release this feature is used without any
|
|
configuration necessary in case the following minimum software versions
|
|
are used:
|
|
|
|
* OVS 2.6
|
|
* DPDK 16.07
|
|
* QEMU 2.7
|
|
|
|
The support of this feature is not yet present in ML2 OVN and ODL
|
|
mechanism drivers.
|
|
|
|
Using vhost-user interfaces
|
|
~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
|
|
|
Once OVS and neutron are correctly configured with DPDK support,
|
|
``vhost-user`` interfaces are completely transparent to the guest
|
|
(except in case of multiqueue configuration described below).
|
|
However, guests must request huge pages. This can be done through flavors.
|
|
For example:
|
|
|
|
.. code-block:: console
|
|
|
|
$ openstack flavor set m1.large --property hw:mem_page_size=large
|
|
|
|
For more information about the syntax for ``hw:mem_page_size``, refer to the
|
|
`Flavors <https://docs.openstack.org/nova/latest/admin/flavors.html>`__ guide.
|
|
|
|
.. note::
|
|
|
|
``vhost-user`` requires file descriptor-backed shared memory. Currently, the
|
|
only way to request this is by requesting large pages. This is why instances
|
|
spawned on hosts with OVS-DPDK must request large pages. The aggregate
|
|
flavor affinity filter can be used to associate flavors with large page
|
|
support to hosts with OVS-DPDK support.
|
|
|
|
Create and add ``vhost-user`` network interfaces to instances in the same
|
|
fashion as conventional interfaces. These interfaces can use the kernel
|
|
``virtio-net`` driver or a DPDK-compatible driver in the guest
|
|
|
|
.. code-block:: console
|
|
|
|
$ openstack server create --nic net-id=$net_id ... testserver
|
|
|
|
Using vhost-user multiqueue
|
|
~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
|
|
|
To use this feature, the following should be set in the flavor extra specs
|
|
(flavor keys):
|
|
|
|
.. code-block:: console
|
|
|
|
$ openstack flavor set $m1.large --property hw:vif_multiqueue_enabled=true
|
|
|
|
This setting can be overridden by the image metadata property if the feature
|
|
is enabled in the extra specs:
|
|
|
|
.. code-block:: console
|
|
|
|
$ openstack image set --property hw_vif_multiqueue_enabled=true IMAGE_NAME
|
|
|
|
Support of ``virtio-net`` multiqueue needs to be present in kernel of
|
|
guest VM and is available starting from Linux kernel 3.8.
|
|
|
|
Check pre-set maximum for number of combined channels in channel
|
|
configuration.
|
|
Configuration of OVS and flavor done successfully should result in
|
|
maximum being more than '1'):
|
|
|
|
.. code-block:: console
|
|
|
|
$ ethtool -l INTERFACE_NAME
|
|
|
|
To increase number of current combined channels run following command in
|
|
guest VM:
|
|
|
|
.. code-block:: console
|
|
|
|
$ ethtool -L INTERFACE_NAME combined QUEUES_NR
|
|
|
|
The number of queues should typically match the number of vCPUs
|
|
defined for the instance. In newer kernel versions
|
|
this is configured automatically.
|
|
|
|
Known limitations
|
|
~~~~~~~~~~~~~~~~~
|
|
|
|
* This feature is only supported when using the libvirt compute driver, and the
|
|
KVM/QEMU hypervisor.
|
|
* Huge pages are required for each instance running on hosts with OVS-DPDK.
|
|
If huge pages are not present in the guest, the interface will appear but
|
|
will not function.
|
|
* Expect performance degradation of services using tap devices: these devices
|
|
do not support DPDK. Example services include DVR and FWaaS.
|
|
* When the ``ovs_use_veth`` option is set to ``True``, any traffic sent
|
|
from a DHCP namespace will have an incorrect TCP checksum.
|
|
This means that if ``enable_isolated_metadata`` is set to ``True`` and
|
|
metadata service is reachable through the DHCP namespace, responses from
|
|
metadata will be dropped due to an invalid checksum. In such cases,
|
|
``ovs_use_veth`` should be switched to ``False`` and Open vSwitch (OVS)
|
|
internal ports should be used instead.
|