Support for virtio-net rx/tx queue sizes¶
With the current multi-queue virtio-net approach, network performance scales as the number of CPUs increase. Large efforts are continuing to increase the performance of network (typically DPDK applications) in order to take advantage of the full capacity of each individual vCPU.
vCPUs can be preempted by the hypervisor kernel thread even with a strong partitioning in place (isolcpus, tuned). The preemption are not frequent, few per seconds, but with 256 descriptor per virtio queue, just one preemption of the vCPU will lead to packet drop, as the 256 slots are filled during the preemption: this is the case for NFV VMs, where the per queue packet rate is above 1 Mpps (1 million of packets per second).
Larger queues sizes permit to amortizing the vCPU preemption and avoiding packet drops, which is one of the main NFV requirements: zero packet drop. With today’s NFV VMs, 1k queues permit achieving high performance without any packet drop, so being able to increase the default queue size to 1k will solves today’s issues. Making it configurable, per operator choice, will also help to address future NFV use cases (we are now shifting from 10Gbps NICs to 25Gbps NICs).
In QEMU 2.7.0 and libvirt 2.3.0 a new tunable has been introduced to configure RX queue size. In QEMU 2.10.0 and libvirt 3.7.0 a new tunable has been introduced to update TX queue size of virtio NICs.
The proposed change is to add new options in “nova.conf” under the “libvirt” section to update the default values of guest XML started on host. So all guests that are using vif type model virtio backed to vhost (VIF_TYPE_BRIDGE, VIF_TYPE_OVS) or backed to vhostuser (VIF_TYPE_VHOSTUSER) will be booted with those new values.
For both options, None will be the default, meaning that libvirt driver will not include any value for the queue size in the domain XML of the guests.
The current implementation allows the size to be between 256 and 1024, and that number should be a power of two. A warning message will be printed for each guest started if the previous constraints are not satisfied.
In case the values are updated while guests are running on the host, only new guests booted will take advantage of those new values. Meaning that the “old” guests running will remaining with the previous values.
NOTE: Two special cases, during hard-reboot and cold-migration, the domain XML of guests are regenerated meaning that they will take advantage of the new values configured in “nova.conf”.
NOTE: Trying to migrate guests which have queue sizes configured to host which have libvirt or QEMU version that does not support such configuration will not work. The migration process will rollback letting the scheduler finding new host.
Operators could achieve this by manually update domain XML of guests.
In Nova, configuring the tx/rx queues could also be achieved by updating the binding profile of the ports created for an instance. But it has been noticed that the tx/rx queues option is not related to Networking but specific to the virtio framework, then allowing users to update these values is not desirable since to correctly configure them the hardware should be knowing.
In Nova, configuring the tx/rx queue sizes could also be achieved by introducing flavor extra-specs which would have provided ability for operators to restrict that feature to certain tenants, flexibility to update the values and manage the hosts using aggregates. Ability to configure different size for different flavors.
Data model impact¶
REST API impact¶
Other end user impact¶
We are assuming that the compute nodes configured with new values for network improvement are going to be isolated in specific aggregates. A user who wants to take advantage of that network improvement would use flavors that are configured to boot on those specifics aggregates.
Other deployer impact¶
Operators must configure tx_queue_size and/or rx_queue_size via “nova.conf”, then isolate the hosts on specific host aggregates and configure flavor properties that will match properties of that aggregate.
Alternatively, operators can avoid isolating hosts.
- Primary assignee:
Sahid Orentino Ferdjaoui <sahid-ferdjaoui>
Introduce new options in nova.conf
Update config xml to handle the new options at boot
Unit tests are going to be added to ensure that the values of tx_queue_size and/or rx_queue_size updated in “nova.conf” will be take into account by guests booted on host.
The documentation of the VIFs that are supporting this feature should be updated as-well as the introduced config options in “nova.conf”.