os-vif/releasenotes/notes/default-qos-policy-for-ovs-...

33 lines
2.0 KiB
YAML

---
upgrade:
- |
A new config option has been added to the OpenvSwitch plugin
``[os_vif_ovs]default_qos_type``. This option controls
the Default tc qdisc applied to a kernel interface attached to OpenvSwitch
on Linux hosts. As of this release, the default tc qdisc is ``linux-noop``
other supported values are ``linux-htb``, ``linux-hfsc``,
``linux-sfq``, ``linux-codel`` and ``linux-fq_codel``.
before this release the default qdisc was undefined. older kernels did not
apply /proc/sys/net/core/default_qdisc to tap devices. newer kernels such
as the one found in rhel 9 do. This can significantly impact performance.
See bug https://bugs.launchpad.net/os-vif/+bug/2017868 for more details.
The default ``linux-noop`` should perform well for all use-cases so no
explicit action is required on upgrade however it should be noted that
the default_qos_type is only set when a port is first created. As such
this fix will not take effect until the next time the vm interface is
recreated. If you change this value for an existing port it will only
take effect after a hard reboot of the VM or a move operation.
fixes:
- |
A significant performance regression was observed on a subset of Linux
kernels and sysctl configurations resulting in a reduction of throughput
to between 10% of the prior performance for small packets and 50% for
large packets. This has now been resolved by setting a default
qos_type on ovs interfaces when they are first created. To mimic libvirt's
undocumented behavior the ``linux-noop`` type is set on the ovs port when
it is first created. This will be overridden by neutron if a qos policy
is defined for a port and is simply the initial value to use when first
adding a port to OpenvSwitch. The default QoS type applied can be
controlled by the ``[os_vif_ovs]default_qos_type`` config operation.
See bug https://bugs.launchpad.net/os-vif/+bug/2017868 for more details.