Using ports with resource request¶
Starting from microversion 2.72 nova supports creating servers with neutron
ports having resource request visible as a admin-only port attribute
resource_request
. For example a neutron port has resource request if it has
a QoS minimum bandwidth rule attached.
The Quality of Service (QoS): Guaranteed Bandwidth document describes how to configure neutron to use this feature.
Resource allocation¶
Nova collects and combines the resource request from each port in a boot request and sends one allocation candidate request to placement during scheduling so placement will make sure that the resource request of the ports are fulfilled. At the end of the scheduling nova allocates one candidate in placement. Therefore the requested resources for each port from a single boot request will be allocated under the server’s allocation in placement.
Resource Group policy¶
Nova represents the resource request of each neutron port as a separate
Granular Resource Request group
when querying placement for allocation candidates. When a server create request
includes more than one port with resource requests then more than one group
will be used in the allocation candidate query. In this case placement requires
to define the group_policy
. Today it is only possible via the
group_policy
key of the flavor extra_spec.
The possible values are isolate
and none
.
When the policy is set to isolate
then each request group and therefore the
resource request of each neutron port will be fulfilled from separate resource
providers. In case of neutron ports with vnic_type=direct
or
vnic_type=macvtap
this means that each port will use a virtual function
from different physical functions.
When the policy is set to none
then the resource request of the neutron
ports can be fulfilled from overlapping resource providers. In case of neutron
ports with vnic_type=direct
or vnic_type=macvtap
this means the ports
may use virtual functions from the same physical function.
For neutron ports with vnic_type=normal
the group policy defines the
collocation policy on OVS bridge level so group_policy=none
is a reasonable
default value in this case.
If the group_policy
is missing from the flavor then the server create
request will fail with ‘No valid host was found’ and a warning describing the
missing policy will be logged.
Virt driver support¶
Supporting neutron ports with vnic_type=direct
or vnic_type=macvtap
depends on the capability of the virt driver. For the supported virt drivers
see the Support matrix
If the virt driver on the compute host does not support the needed capability then the PCI claim will fail on the host and re-schedule will be triggered. It is suggested not to configure bandwidth inventory in the neutron agents on these compute hosts to avoid unnecessary reschedule.
Extended resource request¶
It is expected that neutron 20.0.0 (Yoga) will implement an extended resource
request format via the the port-resource-request-groups
neutron API
extension. As of nova 24.0.0 (Xena), nova already supports this extension if
every nova-compute service is upgraded to Xena version and the
upgrade_levels.compute
configuration does not prevent
the computes from using the latest RPC version.
The extended resource request allows a single Neutron port to request
resources in more than one request groups. This also means that using just one
port in a server create request would require a group policy to be provided
in the flavor. Today the only case when a single port generates more than one
request groups is when that port has QoS policy with both minimum bandwidth
and minimum packet rate rules. Due to the placement resource model of these
features in this case the two request groups will always be fulfilled from
separate resource providers and therefore neither the group_policy=none
nor the group_policy=isolate
flavor extra specs will result in any
additional restriction on the placement of the resources. In the multi port
case the Resource Group policy section above still applies.