OpenStack Compute uses a shared-nothing, messaging-based architecture. While very flexible, the fact that you can install each nova- service on an independent server means there are many possible methods for installing OpenStack Compute. Here are the types of installation architectures:
Single node: Only one server runs all nova- services and also drives all the virtual instances. Use this configuration only for trying out OpenStack Compute, or for development purposes.
Two nodes: A cloud controller node runs the nova- services except for nova-compute, and a compute node runs nova-compute. A client computer is likely needed to bundle images and interfacing to the servers, but a client is not required. Use this configuration for proof of concepts or development environments.
Multiple nodes: You can add more compute nodes to the two node installation by simply installing nova-compute on an additional server and copying a
nova.conffile to the added node. This would result in a multiple node installation. You can also add a volume controller and a network controller as additional nodes in a more complex multiple node installation. A minimum of 4 nodes is best for running multiple virtual instances that require a lot of processing power.
This is an illustration of one possible multiple server installation of OpenStack Compute; virtual server networking in the cluster may vary.
An alternative architecture would be to add more messaging servers
if you notice a lot of back up in the messaging queue causing performance
problems. In that case you would add an additional RabbitMQ server in
addition to or instead of scaling up the database server. Your
installation can run any nova- service on any server as long as the
nova.conf is configured to point to the RabbitMQ
server and the server can send messages to the server.
Multiple installation architectures are possible, here is another example illustration.
