PowerVC high availability and scale architecture

Earlier versions of PowerVC followed the All-in-One (AIO) deployment model that involved installing packages and setting up services corresponding to stateless (OpenStack based services like nova, cinder, etc.) and stateful services (MariaDB, RabbitMQetc) on a single system; thus, not following the distributed approach that OpenStack supports. While All-in-one deployment model keeps PowerVC deployment simple and highly manageable, the model denies PowerVC the scale and high availability that it might achieve with a multinode model.

As you move to larger data centers that consist of hundreds of IBM Power® servers, you need a PowerVC deployment that can cater to your demands of scale and availability.

The multi-node architecture of PowerVC caters primarily to the following needs:
  • Prevent single point of failure - Provide highly available management node, such that failure of one node does not impact virtualization and cloud management of the data center that uses PowerVC.
  • Scalability - Distribute load across multiple nodes to support more scale numbers with the same deployment.
PowerVC 2.0.2 release supports three-node architecture. The multi-node deployment consists of certain aspects to provide a HA solution as explained in the subsequent sections here.
This image shows high multinode architecture and control plane scale and resilience in PowerVCversion 2.0.2.

Multiple nodes – Single deployment

A single HA deployment of Power consists of multiple nodes of PowerVC. Each node has all PowerVC stateless and stateful services running as it is and optionally, services that are related to monitoring tools if the user chooses to deploy them. All nodes of PowerVC services that make this deployment are treated as a single deployment rather than as separate PowerVC entities. All these nodes are tied together to provide a highly available solution that is treated as a single unit, even though the internal solution is distributed into three nodes.

This deployment provides a solution such that most of the services run in Active-Active mode that is monitored by Pacemaker or Corosync. A few services that are listed below run in an Active-Passive mode that is again monitored by Pacemaker.
  • Cinder, nova, neutron health
  • Ceilometer
  • TTV
  • Cinder-volume
This solution, thus, avoids a single point of failure.

VIP for accessing multinode deployment

Even though the different PowerVC services run running across three different nodes, in HA configuration, the user remains oblivious to this. The PowerVC HA deployment is accessible only through a virtual IP address (VIP) that internally routes the request to one of the three PowerVC management nodes with the help of a load-balancer (HAProxy). This also means that managed nodes like the NovaLink compute nodes access PowerVC through VIP and not through the individual node IP.