One common question from OpenStack operators is that “how does the control plane (e.g., database, messaging queue, nova-scheduler ) scales?”. To answer this question, operators setup Rally to drive workload to the OpenStack cloud. However, without a large number of nova-compute nodes, it becomes difficult to exercise the control performance.
Given the built-in feature of Docker container, Kolla enables standing up many of nova-compute nodes with nova fake driver on a single host. For example, we can create 100 nova-compute containers on a real host to simulate the 100-hypervisor workload to the nova-conductor and the messaging queue.
Nova fake driver can not work with all-in-one deployment. This is because the fake neutron-openvswitch-agent for the fake nova-compute container conflicts with neutron-openvswitch-agent on the compute nodes. Therefore, in the inventory the network node must be different than the compute node.
By default, Kolla uses libvirt driver on the compute node. To use nova-fake
driver, edit the following parameters in ansible/group_vars
or in the
command line options.
enable_nova_fake: "yes"
num_nova_fake_per_node: 5
Each compute node will run 5 nova-compute containers and 5 neutron-plugin-agent containers. When booting instance, there will be no real instances created. But nova list shows the fake instances.
Except where otherwise noted, this document is licensed under Creative Commons Attribution 3.0 License. See all OpenStack Legal Documents.