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.
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 website.
OVS with DPDK, or OVS-DPDK, can be used to provide high-performance networking between instances on OpenStack compute nodes.
Using DPDK in OVS requires the following minimum software versions:
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:
In both cases, install and configure Open vSwitch with DPDK support for each node. For more information, see the OVS-DPDK installation guide (select an appropriate OVS version in the Branch drop-down menu).
Neutron configuration reference for OVS-DPDK for configuration of neutron OVS agent.
In case you wish to configure multiqueue, see the OVS configuration chapter on vhost-user in QEMU documentation.
The technical background of multiqueue is explained in the corresponding blueprint.
Additionally, OpenStack supports vhost-user
reconnect feature starting
from the Ocata release, as implementation of fix for
bug 1604924.
Starting from OpenStack Ocata release this feature is used without any
configuration necessary in case the following minimum software versions
are used:
The support of this feature is not yet present in ML2 OVN and ODL mechanism drivers.
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:
$ 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 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
$ openstack server create --nic net-id=$net_id ... testserver
To use this feature, the following should be set in the flavor extra specs (flavor keys):
$ 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:
$ openstack image set --property hw_vif_mutliqueue_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’):
$ ethtool -l INTERFACE_NAME
To increase number of current combined channels run following command in guest VM:
$ 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.
Except where otherwise noted, this document is licensed under Creative Commons Attribution 3.0 License. See all OpenStack Legal Documents.