Configuring Receive-Side Scaling (RSS) in OpenStack

Enabling the multiqueue support allows the guest instances to increase the network throughput.

1. Edit the nova.conf file on the target compute node.

   vim /etc/nova/nova.conf

   Find the [PCI] section. The following shows an example for a VLAN-tunneled network.

   [pci]
   passthrough_whitelist = {"address":"0000:3b:00.4","physical_network":"public"}
   passthrough_whitelist = {"address":"0000:3b:00.5","physical_network":"public"}
   passthrough_whitelist = {"address":"0000:3b:00.6","physical_network":"public"}

   Set the number of queues as follows.

   [pci]
   passthrough_whitelist = {"address":"0000:3b:00.4","physical_network":"public","queues":2}
   passthrough_whitelist = {"address":"0000:3b:00.5","physical_network":"public","queues":2}
   passthrough_whitelist = {"address":"0000:3b:00.6","physical_network":"public","queues":2}

   An example for a VXLAN-tunneled network:

   [pci]
   passthrough_whitelist = {"address":"0000:3b:00.04"}
   passthrough_whitelist = {"address":"0000:3b:00.05"} 
   passthrough_whitelist = {"address":"0000:3b:00.06"}

   Set the number of queues as follows.

   [pci]
   passthrough_whitelist = {"address":"0000:3b:00.04","queues":2}
   passthrough_whitelist = {"address":"0000:3b:00.05","queues":2} 
   passthrough_whitelist = {"address":"0000:3b:00.06","queues":2}

2. Restart the nova services on the compute node.

   systemctl restart openstack-nova-*

3. On the controller node, create a VM with the target number of queues.

   openstack flavor create --public <flavor_name> --id auto \
   --ram <RAM_size_in_MB> --disk <disk_size_in_GB> --vcpus <number_of_vCPU>

   For example:

   openstack flavor create --public p1.medium --id auto \
   --ram 4096 --disk 40 --vcpus 4

   Set the number of queues for a port.

   openstack port set --binding-profile queues=<number_of_queues> <port_name>

   For example:

   openstack port set --binding-profile queues=2 port10

   --binding-profile queues=2 instructs the OpenStack scheduler to find a virtual function with the specified number of queues. If a virtual function with the specified number of queues is not found in any passthrough_whitelist, or virtual functions are not available, the VM will not start.

   Note: If the queue parameter is not specified with the port set --binding-profile command, the OpenStack scheduler will take any first available virtual function from passthrough_whitelist regardless of the number of queues defined in the passthrough_whitelist configuration. It is therefore recommended to create a port and a VM with the target number of queues using the queue parameter in the port set --binding-profile command.

   Enable huge pages for the flavor.

   openstack flavor set <flavor_name> --property hw:mem_page_size=large

   For example:

   openstack flavor set p1.medium --property hw:mem_page_size=large

   Virtio packed ring can be used if required. Set virtio packed ring as follows.

   openstack image set fedora --property driver_packed='True'

   After the flavor configuration, a VM instance can be created.

   openstack server create --flavor <flavor_name> \
   --image <image_name> --port <port_name> <vm_name>

   For example:

   openstack server create --flavor m1.medium \
   --image fedora_cloud_qcow --port port10 vm1