The best-effort pod should be immediately dispatched in allocate function of session , without considering the gang constraint

[sivanzcw]

/kind bug

Considering the gang constraint in the backfill action may cause the job with gang configuration and contains the best-effort pod to be unscheduled.

In backfill action, if task predicated succeed, the scheduler will call ssn.Allocate function to dispatch pod

for _, node := range ssn.Nodes {
	// TODO (k82cn): predicates did not consider pod number for now, there'll
	// be ping-pong case here.
	if err := ssn.PredicateFn(task, node); err != nil {
		klog.V(3).Infof("Predicates failed for task <%s/%s> on node <%s>: %v",
			task.Namespace, task.Name, node.Name, err)
		fe.SetNodeError(node.Name, err)
		continue
	}

	klog.V(3).Infof("Binding Task <%v/%v> to node <%v>", task.Namespace, task.Name, node.Name)
	if err := ssn.Allocate(task, node.Name); err != nil {
		klog.Errorf("Failed to bind Task %v on %v in Session %v", task.UID, node.Name, ssn.UID)
		fe.SetNodeError(node.Name, err)
		continue
	}

	allocated = true
	break
}

In ssn.Allocate function, only when job of task is ready, can the task be dispatched

if ssn.JobReady(job) {
	for _, task := range job.TaskStatusIndex[api.Allocated] {
		if err := ssn.dispatch(task); err != nil {
			klog.Errorf("Failed to dispatch task <%v/%v>: %v",
				task.Namespace, task.Name, err)
			return err
		}
	}
}

If others pods under the job are not schedulerd, the best-effort pod can not be scheduled. If all pods under the job are best-effort, the job will not be scheduled.

In addition, for best-effort pods, there are no resource requests, it it not necessary to consider the gang constraint.

[carmark]

This may make the MPIJob pending when the launcher does not config resource requests/limits.

[k82cn]

/cc @carmark

[sivanzcw]

closed by #647