The number of compute nodes in a cluster can auto-scale depending on the job
queue depth. This behavior is controlled by the Jobs per Compute Node
setting and constrained by the Minimum Nodes
and Maximum Nodes
settings.
A cluster will add a new compute node if the following condition is met:
node_count < floor(queue_depth / jobs_per_compute_node)
A cluster will remove a compute node if the following condition is met:
node_count > ceil(queue_depth / jobs_per_compute_node)
If jobs_per_compute_node = 1000
and the queue starts empty, the cluster will
use a single compute node until queue_depth
reaches 2000, at which point it
will add a second compute node.
As the queue shrinks, the cluster will attempt to remove a compute node when
queue_depth
drops below 1000.
The cluster will attempt to minimize your costs by never removing a compute node that is processing a job, and only removing compute nodes that are near the end of a billed hour.