5. Scalability

At some point, you might want to increase the size of the cluster or achieve high availability (also refered as HA), or just allow some kind of load balancing.

I propose a strategy based on my personal experience, in hope it might suits your needs.

The concept is based on the Corrosync / Pacemaker couple, along few other tools.

Architecture:

• X worker nodes, for example 2000.

These X nodes are split into “herds”, which are logical groups of nodes that a single management node is able to handle. 500 or 600 seems a reasonable size for an herd.

• N management nodes, N being equal to the number “herds”. In our example, 2000 / 500 = 4, so 4 management nodes.

• 1 NFS storage bay, with hardware redundancy. This component is critical and should be chosen with care. It will be mounted on each management node.

  Note

  It is possible to get ride of the NFS storage bay, using replication tools like DRBD or any other parallel FS.