In metric-based dispatch mode, the Load Balancer evaluates nodes in a round-robin fashion until it finds a node that can accept the task. The Load Balancer checks the resource provision thresholds on the first available node. It dispatches the task to this node if dispatching the task causes none of the thresholds to be exceeded. If dispatching the task causes any threshold to be exceeded, or if the node is out of free swap space, the Load Balancer evaluates the next node. It continues to evaluate nodes until it finds a node that can accept the task.
To determine whether a task can run on a particular node, the Load Balancer collects and stores statistics from the last three runs of the task. It compares these statistics with the resource provision thresholds defined for the node. If no statistics exist in the repository, the Load Balancer uses the following default values:
40 MB memory
The Load Balancer dispatches tasks for execution in the order the Workflow Manager or scheduler submits them. The Load Balancer does not bypass any tasks in the dispatch queue. Therefore, if a resource intensive task is first in the dispatch queue, all other tasks with the same service level must wait in the queue until the Load Balancer dispatches the resource intensive task.