RX IRQs triggering for network traffic processing in multiqueue-enabled virtual machines

• Setup a QEMU/KVM environment [1];
• Install openvswitch supporting dpdk (>= 2.14.2);
• On openvswitch, configure a bridge including two dpdkvhostuserclient interfaces as described in [2], enabling multiqueue support on them;
• Configure two SLES VMs, each one having a vNIC attached to the bridge described above and setup with multiple queues in a virtio-net / vhostuser configuration [3];

The number of RX IRQs being triggered is related to the number of Poll Mod Drivers (PMDs) being run by Openvswitch.
Their configuration can be shown by running the command :

ovs-appctl dpif-netdev/pmd-rxq-show

To fine tune them, the pmd-cpu-mask configuration parameter is available, which specifies the CPU mask for setting the CPU affinity of the PMD threads.

The value is specified in hexadecimal (hex) format: when not specified, only one PMD thread is created per each NUMA node and, by default, pinned to any available CPU on the NUMA node.

For example, the following command assigns CPUs 2, 5, and 8 to PMDs handling:

ovs-vsctl set Open_vSwitch . other_config:pmd-cpu-mask=" 0x124"

The wider the bitmask, the more CPUs dedicated to PMDs, the more distinct RX IRQs can be triggered in the VMs.

Moreover, it is worth explicitly remarking that the RX queues being assigned to non-dedicated PMDs will not be polled.

[1] https://documentation.suse.com/sles/15-SP5/html/SLES-all/cha-qemu-host.html#
[2] https://docs.openvswitch.org/en/latest/topics/dpdk/vhost-user/
[3] https://libvirt.org/formatdomain.html#vhost-user-interface