Quality of Service (QoS): Guaranteed Minimum
Bandwidth
Most Networking Quality of Service (QoS) features are implemented
solely by OpenStack Neutron and they are already documented in the QoS
.
configuration chapter of the Networking Guide <config-qos>
Some more complex QoS features necessarily involve the scheduling of a
cloud server, therefore their implementation is shared between OpenStack
Nova, Neutron and Placement. As of the OpenStack Stein release the
Guaranteed Minimum Bandwidth feature is like the latter.
This Networking Guide chapter does not aim to replace Nova or
Placement documentation in any way, but it still hopes to give an
overall OpenStack-level guide to understanding and configuring a
deployment to use the Guaranteed Minimum Bandwidth feature.
A guarantee of minimum available bandwidth can be enforced on two
levels:
- Scheduling a server on a compute host where the bandwidth is
available. To be more precise: scheduling one or more ports of a server
on a compute host’s physical network interfaces where the bandwidth is
available. - Queueing network packets on a physical network interface to provide
the guaranteed bandwidth.
In short the enforcement has two levels:
- (server) placement and
- data plane.
Since the data plane enforcement is already documented in the QoS chapter <config-qos>
, here we only document
the placement-level enforcement.
Limitations
- A pre-created port with a
minimum-bandwidth
rule must
be passed when booting a server (openstack server create
).
Passing a network with a minimum-bandwidth rule at boot is not supported
because of technical reasons (in this case the port is created too late
for Neutron to affect scheduling). - Bandwidth guarantees for ports can only be requested on networks
backed by a physical network (physnet). - In Stein there is no support for networks with multiple physnets.
However some simpler multi-segment networks are still supported:- Networks with multiple segments all having the same physnet
name. - Networks with only one physnet segment (the other segments being
tunneled segments).
- Networks with multiple segments all having the same physnet
- If you mix ports with and without bandwidth guarantees on the same
physical interface then the ports without a guarantee may starve.
Therefore mixing them is not recommended. Instead it is recommended to
separate them byNova host aggregates <admin/aggregates>
. - Changing the guarantee of a QoS policy (adding/deleting a
minimum_bandwidth
rule, or changing the
min_kbps
field of aminimum_bandwidth
rule) is
only possible while the policy is not in effect. That is ports of the
QoS policy are not yet used by Nova. Requests to change guarantees of
in-use policies are rejected. - Changing the QoS policy of the port with new
minimum_bandwidth
rules changes placement
allocations
from Wallaby release. If the VM was booted with
port without QoS policy andminimum_bandwidth
rules the
port update succeeds but placement allocations will not change. The same
is true if the port has nobinding:profile
, thus no
placement allocation record exists for it. But if the VM was booted with
a port with QoS policy andminimum_bandwidth
rules the
update is possible and the allocations are changed in placement as
well.
Note
As it is possible to update a port to remove the QoS policy, updating
it back to have QoS policy with minimum_bandwidth
rule will
not result in placement allocation
record, only the
dataplane enforcement will happen.
Note
updating the minimum_bandwidth
rule of a QoS policy that
is attached to a port which is bound to a VM is still not possible.
- The first data-plane-only Guaranteed Minimum Bandwidth
implementation (for SR-IOV egress traffic) was released in the Newton
release of Neutron. Because of the known lack of placement-level
enforcement it was marked as “best
effort” (5th bullet point). Since placement-level enforcement was
not implemented bandwidth may have become overallocated and the system
level resource inventory may have become inconsistent. Therefore for
users of the data-plane-only implementation a migration/healing process
is mandatory (see section On
Healing of Allocations) to bring the system level resource inventory
to a consistent state. Further operations that would reintroduce
inconsistency (e.g. migrating a server with
minimum_bandwidth
QoS rule, but no resource allocation in
Placement) are rejected now in a backward-incompatible way. - The Guaranteed Minimum Bandwidth feature is not complete in the
Stein release. Not all Nova server lifecycle operations can be executed
on a server with bandwidth guarantees. Since Stein (Nova API
microversion 2.72+) you can boot and delete a server with a guarantee
and detach a port with a guarantee. Since Train you can also migrate and
resize a server with a guarantee. Support for further server move
operations (for example evacuate, live-migrate and unshelve after
shelve-offload) is to be implemented later. For the definitive
documentation please refer to the Port
with Resource Request chapter of the OpenStack Compute API
Guide. - If an SR-IOV physical function is configured for use by the
neutron-openvswitch-agent, and the same physical function’s virtual
functions are configured for use by the neutron-sriov-agent then the
available bandwidth must be statically split between the corresponding
resource providers by administrative choice. For example a 10 Gbps
SR-IOV capable physical NIC could be treated as two independent NICs -a
5 Gbps NIC (technically the physical function of the NIC) added to an
Open vSwitch bridge, and another 5 Gbps NIC whose virtual functions can
be handed out to servers by neutron-sriov-agent. - Neutron allows physnet names to be case sensitive. So physnet0 and
Physnet0 are treated as different physnets. Physnets are mapped to
traits in Placement for scheduling purposes. However Placement traits
are case insensitive and normalized to full capital. Therefore the
scheduling treats physnet0 and Physnet0 as the same physnet. It is
advised not to use physnet names that are only differ by case. - There are hardware platforms (e.g.: Cavium ThunderX) where it’s
possible to have virtual functions which are network devices that are
not associated to a physical function. As bandwidth resources are
tracked per physical function, for such hardware the placement
enforcement of the QoS minimum bandwidth rules cannot be supported.
Creating a server with ports using such QoS policy targeting such
hardware backend will result in aNoValidHost
error during
scheduling. - When QoS is used with a trunk, Placement enforcement is applied only
to the trunk’s parent port. Subports are not going to have Placement
allocation. As a workaround, parent port’s QoS policy should take into
account subports needs and request enough minimum bandwidth resources to
accommodate every port in the trunk.
Placement pre-requisites
Placement must support microversion
1.29. This was first released in Rocky.
Nova pre-requisites
Nova must support microversion
2.72. This was first released in Stein.
Not all Nova virt drivers are supported, please refer to the Virt
Driver Support section of the Nova Admin Guide.
Neutron pre-requisites
Neutron must support the following API extensions:
agent-resources-synced
port-resource-request
qos-bw-minimum-ingress
These were all first released in Stein.
Supported drivers and agents
In release Stein the following agent-based ML2 mechanism drivers are
supported:
- Open vSwitch (
openvswitch
) vnic_types:
normal
,direct
- SR-IOV (
sriovnicswitch
) vnic_types:
direct
,macvtap
,
direct-physical
- OVN (
ovn
) vnic_types:normal
Note
SR-IOV (sriovnicswitch
) agent does not handle
direct-physical
ports. However the agent can report the
bandwidth capacity of a network device that will be used by a
direct-physical
port.
neutron-server config
The placement
service plugin synchronizes the agents’
resource provider information from neutron-server to Placement.
Since neutron-server talks to Placement you need to configure how
neutron-server should find Placement and authenticate to it.
/etc/neutron/neutron.conf
(on controller nodes):
[DEFAULT]
service_plugins = placement,...
auth_strategy = keystone
[placement]
auth_type = password
auth_url = https://controller/identity
password = secret
project_domain_name = Default
project_name = service
user_domain_name = Default
username = placement
If a vnic_type is supported by default by multiple ML2 mechanism
drivers (e.g. vnic_type=direct
by both
openvswitch
and sriovnicswitch
) and multiple
agents’ resources are also meant to be tracked by Placement, then the
admin must decide which driver to take ports of that vnic_type by
prohibiting the vnic_type for the unwanted drivers. Use ovs_driver.vnic_type_prohibit_list
in
this case. Valid values are all the supported_vnic_types
of
the respective
mechanism drivers.
/etc/neutron/plugins/ml2/ml2_conf.ini
(on controller
nodes):
neutron-openvswitch-agent
config
Set the agent configuration as the authentic source of the resources
available. Set it on a per-bridge basis by ovs.resource_provider_bandwidths
. The
format is: bridge:egress:ingress,...
You may set only one
direction and omit the other.
Note
egress
/ ingress
is meant from the
perspective of a cloud server. That is egress
= cloud
server upload, ingress
= download.
Egress and ingress available bandwidth values are in
kilobit/sec (kbps)
.
If desired, resource provider inventory fields can be tweaked on a
per-agent basis by setting ovs.resource_provider_inventory_defaults
.
Valid values are all the optional
parameters of the update resource provider inventory call.
/etc/neutron/plugins/ml2/ovs_agent.ini
(on compute and
network nodes):
[ovs]
bridge_mappings = physnet0:br-physnet0,...
resource_provider_bandwidths = br-physnet0:10000000:10000000,...
#resource_provider_inventory_defaults = step_size:1000,...
neutron-sriov-agent config
The configuration of neutron-sriov-agent is analog to that of
neutron-openvswitch-agent. However look out for:
- The different .ini section names as you can see below.
- That neutron-sriov-agent allows a physnet to be backed by multiple
physical devices. - Of course refer to SR-IOV physical functions instead of bridges in
sriov_nic.resource_provider_bandwidths
.
/etc/neutron/plugins/ml2/sriov_agent.ini
(on compute
nodes):
[sriov_nic]
physical_device_mappings = physnet0:ens5,physnet0:ens6,...
resource_provider_bandwidths = ens5:40000000:40000000,ens6:40000000:40000000,...
#resource_provider_inventory_defaults = step_size:1000,...
OVN chassis config
Bandwidth config values are stored in each SB chassis register, in
“external_ids:ovn-cms-options”. The configuration options are the same
as in SR-IOV and OVS agents. This is how the values are registered:
$ root@dev20:~# ovs-vsctl list Open_vSwitch
...
external_ids : {hostname=dev20.fistro.com, \
ovn-cms-options="resource_provider_bandwidths=br-ex:1001:2000;br-ex2:3000:4000, \
resource_provider_inventory_defaults=allocation_ratio:1.0;min_unit:10, \
resource_provider_hypervisors=br-ex:dev20.fistro.com;br-ex2:dev20.fistro.com", \
rundir="/var/run/openvswitch", \
system-id="029e7d3d-d2ab-4f2c-bc92-ec58c94a8fc1"}
...
Each configuration option defined in “external_ids:ovn-cms-options”
is divided by commas.
This information is retrieved from the OVN SB database during the
Neutron server initialization and when the “Chassis” registers are
updated.
During the Neutron server initialization, a
MaintenanceWorker
thread will call
OvnSbSynchronizer.do_sync
, that will call
OVNClientPlacementExtension.read_initial_chassis_config
.
This method lists all chassis and builds the resource provider
information needed by Placement. This information is stored in the
“Chassis” registers, in “external_ids:ovn-cms-options”, with the same
format as retrieved from the local “Open_vSwitch” registers from each
chassis.
The second method to update the Placement information is when a
“Chassis” registers is updated. The
OVNClientPlacementExtension
extension registers an event
handler that attends the OVN SB “Chassis” bandwidth configuration
changes. This event handler builds a PlacementState
instance and sends it to the Placement API. If a new chassis is added or
an existing one changes its resource provider configuration, this event
updates it in the Placement database.
Propagation of resource
information
The flow of information is different for available and used
resources.
The authentic source of available resources is neutron agent
configuration -where the resources actually exist, as described in the
agent configuration sections above. This information is propagated in
the following chain:
neutron-l2-agent -> neutron-server -> Placement
.
From neutron agent to server the information is included in the
configurations
field of the agent heartbeat message sent on
the message queue periodically.
# as admin
$ openstack network agent list --agent-type open-vswitch --host devstack0
+--------------------------------------+--------------------+-----------+-------------------+-------+-------+---------------------------+
| ID | Agent Type | Host | Availability Zone | Alive | State | Binary |
+--------------------------------------+--------------------+-----------+-------------------+-------+-------+---------------------------+
| 5e57b85f-b017-419a-8745-9c406e149f9e | Open vSwitch agent | devstack0 | None | :-) | UP | neutron-openvswitch-agent |
+--------------------------------------+--------------------+-----------+-------------------+-------+-------+---------------------------+
# output shortened and pretty printed
# note: 'configurations' on the wire, but 'configuration' in the cli
$ openstack network agent show -f value -c configuration 5e57b85f-b017-419a-8745-9c406e149f9e
{'bridge_mappings': {'physnet0': 'br-physnet0'},
'resource_provider_bandwidths': {'br-physnet0': {'egress': 10000000,
'ingress': 10000000}},
'resource_provider_inventory_defaults': {'allocation_ratio': 1.0,
'min_unit': 1,
'reserved': 0,
'step_size': 1},
...
}
Re-reading the resource related subset of configuration on
SIGHUP
is not implemented. The agent must be restarted to
pick up and send changed configuration.
Neutron-server propagates the information further to Placement for
the resources of each agent via Placement’s HTTP REST API. To avoid
overloading Placement this synchronization generally does not happen on
every received heartbeat message. Instead the re-synchronization of the
resources of one agent is triggered by:
- The creation of a network agent record (as queried by
openstack network agent list
). Please note that deleting an
agent record and letting the next heartbeat to re-create it can be used
to trigger synchronization without restarting an agent. - The restart of that agent (technically
start_flag
being
present in the heartbeat message).
Both of these can be used by an admin to force a re-sync if
needed.
The success of a synchronization attempt from neutron-server to
Placement is persisted into the relevant agent’s
resources_synced
attribute. For example:
# as admin
$ openstack network agent show -f value -c resources_synced 5e57b85f-b017-419a-8745-9c406e149f9e
True
resources_synced
may take the value True, False and
None:
- None: No sync was attempted (normal for agents not reporting
Placement-backed resources). - True: The last sync attempt was completely successful.
- False: The last sync attempt was partially or utterly
unsuccessful.
In case resources_synced
is not True for an agent,
neutron-server does try to re-sync on receiving every heartbeat message
from that agent. Therefore it should be able to recover from transient
errors of Neutron-Placement communication (e.g. Placement being started
later than Neutron).
It is important to note that the restart of neutron-server does not
trigger any kind of re-sync to Placement (to avoid an update storm).
As mentioned before, the information flow for resources requested and
(if proper) allocated is different. It involves a conversation between
Nova, Neutron and Placement.
- Neutron exposes a port’s resource needs in terms of resource classes
and traits as the admin-onlyresource_request
attribute of
that port. - Nova reads this and incorporates
it as a numbered request group into the cloud servers overall
allocation candidate request to Placement. - Nova selects (schedules) and allocates one candidate returned by
Placement. - Nova informs Neutron when binding the port of which physical network
interface resource provider had been selected for the port’s resource
request in thebinding:profile.allocation
sub-attribute of
that port.
For details please see slides
13-15 of a (pre-release) demo that was presented on the Berlin
Summit in November 2018.
Since Yoga, the resource_request
attribute of the port
changed. With the extension port-resource-request-groups
,
Neutron informs that the blob passed to Nova can contain several
bandwidth requests. Please check resource_request
sanitization.
Sample usage
Physnets and QoS policies (together with their rules) are usually
pre-created by a cloud admin:
# as admin
$ openstack network create net0 \
--provider-network-type vlan \
--provider-physical-network physnet0 \
--provider-segment 100
$ openstack subnet create subnet0 \
--network net0 \
--subnet-range 10.0.4.0/24
$ openstack network qos policy create policy0
$ openstack network qos rule create policy0 \
--type minimum-bandwidth \
--min-kbps 1000000 \
--egress
$ openstack network qos rule create policy0 \
--type minimum-bandwidth \
--min-kbps 1000000 \
--ingress
Then a normal user can use the pre-created policy to create ports and
boot servers with those ports:
# as an unprivileged user
# an ordinary soft-switched port: ``--vnic-type normal`` is the default
$ openstack port create port-normal-qos \
--network net0 \
--qos-policy policy0
# alternatively an SR-IOV port, unused in this example
$ openstack port create port-direct-qos \
--network net0 \
--vnic-type direct \
--qos-policy policy0
$ openstack server create server0 \
--flavor cirros256 \
--image cirros-0.5.1-x86_64-disk \
--port port-normal-qos
On Healing of Allocations
Since Placement carries a global view of a cloud deployment’s
resources (what is available, what is used) it may in some conditions
get out of sync with reality.
One important case is when the data-plane-only Minimum Guaranteed
Bandwidth feature was used before Stein (first released in Newton).
Since before Stein guarantees were not enforced during server placement
the available resources may have become overallocated without notice. In
this case Placement’s view and the reality of resource usage should be
made consistent during/after an upgrade to Stein.
Another case stems from OpenStack not having distributed transactions
to allocate resources provided by multiple OpenStack components (here
Nova and Neutron). There are known race conditions in which Placement’s
view may get out of sync with reality. The design knowingly minimizes
the race condition windows, but there are known problems:
- If a QoS policy is modified after Nova read a port’s
resource_request
but before the port is bound its state
before the modification will be applied. - If a bound port with a resource allocation is deleted. The port’s
allocation is leaked. https://bugs.launchpad.net/nova/+bug/1820588
Note
Deleting a bound port has no known use case. Please consider
detaching the interface first by
openstack server remove port
instead.
Incorrect allocations may be fixed by:
- Moving the server, which will delete the wrong allocation and create
the correct allocation as soon as move operations are implemented (not
in Stein unfortunately). Moving servers fixes local
overallocations. - The need for an upgrade-helper allocation healing tool is being
tracked in bug
1819923. - Manually, by using openstack
resource provider allocation set /delete.
Debugging
- Are all components running at least the Stein release?
- Is the
placement
service plugin enabled in
neutron-server? - Is
resource_provider_bandwidths
configured for the
relevant neutron agent? - Is
resource_provider_bandwidths
aligned with
bridge_mappings
or
physical_device_mappings
? - Was the agent restarted since changing the configuration file?
- Is
resource_provider_bandwidths
reaching
neutron-server?
# as admin
$ openstack network agent show ... | grep configurations
Please find an example in section Propagation of resource
information.
- Did neutron-server successfully sync to Placement?
# as admin
$ openstack network agent show ... | grep resources_synced
Please find an example in section Propagation of resource
information.
- Is the resource provider tree correct? Is the root a compute host?
One level below the agents? Two levels below the physical network
interfaces?
$ openstack --os-placement-api-version 1.17 resource provider list
+--------------------------------------+------------------------------------------+------------+--------------------------------------+--------------------------------------+
| uuid | name | generation | root_provider_uuid | parent_provider_uuid |
+--------------------------------------+------------------------------------------+------------+--------------------------------------+--------------------------------------+
| 3b36d91e-bf60-460f-b1f8-3322dee5cdfd | devstack0 | 2 | 3b36d91e-bf60-460f-b1f8-3322dee5cdfd | None |
| 4a8a819d-61f9-5822-8c5c-3e9c7cb942d6 | devstack0:NIC Switch agent | 0 | 3b36d91e-bf60-460f-b1f8-3322dee5cdfd | 3b36d91e-bf60-460f-b1f8-3322dee5cdfd |
| 1c7e83f0-108d-5c35-ada7-7ebebbe43aad | devstack0:NIC Switch agent:ens5 | 2 | 3b36d91e-bf60-460f-b1f8-3322dee5cdfd | 4a8a819d-61f9-5822-8c5c-3e9c7cb942d6 |
| 89ca1421-5117-5348-acab-6d0e2054239c | devstack0:Open vSwitch agent | 0 | 3b36d91e-bf60-460f-b1f8-3322dee5cdfd | 3b36d91e-bf60-460f-b1f8-3322dee5cdfd |
| f9c9ce07-679d-5d72-ac5f-31720811629a | devstack0:Open vSwitch agent:br-physnet0 | 2 | 3b36d91e-bf60-460f-b1f8-3322dee5cdfd | 89ca1421-5117-5348-acab-6d0e2054239c |
+--------------------------------------+------------------------------------------+------------+--------------------------------------+--------------------------------------+
- Does Placement have the expected traits?
# as admin
$ openstack --os-placement-api-version 1.17 trait list | awk '/CUSTOM_/ { print $2 }' | sort
CUSTOM_PHYSNET_PHYSNET0
CUSTOM_VNIC_TYPE_DIRECT
CUSTOM_VNIC_TYPE_DIRECT_PHYSICAL
CUSTOM_VNIC_TYPE_MACVTAP
CUSTOM_VNIC_TYPE_NORMAL
- Do the physical network interface resource providers have the proper
trait associations and inventories?
# as admin
$ openstack --os-placement-api-version 1.17 resource provider trait list RP-UUID
$ openstack --os-placement-api-version 1.17 resource provider inventory list RP-UUID
- Does the QoS policy have a
minimum-bandwidth
rule? - Does the port have the proper policy?
- Does the port have a
resource_request
?
# as admin
$ openstack port show port-normal-qos | grep resource_request
- Was the server booted with a port (as opposed to a network)?
- Did nova allocate resources for the server in Placement?
# as admin
$ openstack --os-placement-api-version 1.17 resource provider allocation show SERVER-UUID
- Does the allocation have a part on the expected physical network
interface resource provider?
# as admin
$ openstack --os-placement-api-version 1.17 resource provider show --allocations RP-UUID
- Did placement manage to produce an allocation candidate list to nova
during scheduling? - Did nova manage to schedule the server?
- Did nova tell neutron which physical network interface resource
provider was allocated to satisfy the bandwidth request?
# as admin
$ openstack port show port-normal-qos | grep binding.profile.*allocation
- Did neutron manage to bind the port?
Links
- Pre-release feature
demo presented on the Berlin Summit in November 2018 - Nova documentation on using a port with
resource_request
- Neutron spec: QoS minimum bandwidth allocation in Placement API
- Nova spec: Network Bandwidth resource provider
- Nova spec: QoS minimum guaranteed packet rate
- Relevant OpenStack Networking API references
- Microversion histories
- Implementation
- Known Bugs