fix the proportion plugin that ignore the inqueue resource in running…

… jobs when podgroup's minResource is not nil

Signed-off-by: Thor-wl <13164644535@163.com>

pkg/scheduler/plugins/overcommit/overcommit.go

@@ -88,10 +88,24 @@ func (op *overcommitPlugin) OnSessionOpen(ssn *framework.Session) {
 	}
 	op.idleResource = total.Clone().Multi(op.overCommitFactor).Sub(used)
 
-	// calculate inqueue job resources
+	klog.V(4).Infof("op.inqueueResource.cpu: %f", op.inqueueResource.MilliCPU)
+
 	for _, job := range ssn.Jobs {
+		// calculate inqueue job resources
 		if job.PodGroup.Status.Phase == scheduling.PodGroupInqueue && job.PodGroup.Spec.MinResources != nil {
 			op.inqueueResource.Add(api.NewResource(*job.PodGroup.Spec.MinResources))
+			continue
+		}
+		// calculate inqueue resource for running jobs
+		// the judgement 'job.PodGroup.Status.Running >= job.PodGroup.Spec.MinMember' will work on cases such as the following condition:
+		// Considering a Spark job is completed(driver pod is completed) while the podgroup keeps running, the allocated resource will be reserved again if without the judgement.
+		if job.PodGroup.Status.Phase == scheduling.PodGroupRunning &&
+			job.PodGroup.Spec.MinResources != nil &&
+			job.PodGroup.Status.Running >= job.PodGroup.Spec.MinMember {
+			allocated := util.GetAllocatedResource(job)
+			inqueued := util.GetInqueueResource(job, allocated)
+			klog.V(4).Infof("reserve resource for running job %s, inqueue: %s", job.Name, *inqueued)
+			op.inqueueResource.Add(inqueued)
 		}
 	}
 

pkg/scheduler/plugins/proportion/proportion.go

@@ -140,6 +140,13 @@ func (pp *proportionPlugin) OnSessionOpen(ssn *framework.Session) {
 		if job.PodGroup.Status.Phase == scheduling.PodGroupInqueue {
 			attr.inqueue.Add(job.GetMinResources())
 		}
+		if job.PodGroup.Status.Phase == scheduling.PodGroupRunning &&
+			job.PodGroup.Spec.MinResources != nil &&
+			job.PodGroup.Status.Running >= job.PodGroup.Spec.MinMember {
+			allocated := util.GetAllocatedResource(job)
+			inqueued := util.GetInqueueResource(job, allocated)
+			attr.inqueue.Add(inqueued)
+		}
 	}
 
 	// Record metrics

pkg/scheduler/plugins/util/util.go

@@ -298,3 +298,48 @@ func NormalizeScore(maxPriority int64, reverse bool, scores []api.ScoredNode) {
 		scores[idx].Score = score
 	}
 }
+
+// GetAllocatedResource returns allocated resource for given job
+func GetAllocatedResource(job *api.JobInfo) *api.Resource {
+	allocated := &api.Resource{}
+	for status, tasks := range job.TaskStatusIndex {
+		if api.AllocatedStatus(status) {
+			for _, t := range tasks {
+				allocated.Add(t.Resreq)
+			}
+		}
+	}
+	return allocated
+}
+
+// GetInqueueResource returns reserved resource for running job whose part of pods have not been allocated resource.
+func GetInqueueResource(job *api.JobInfo, allocated *api.Resource) *api.Resource {
+	inqueue := &api.Resource{}
+	for rName, rQuantity := range *job.PodGroup.Spec.MinResources {
+		switch rName {
+		case v1.ResourceCPU:
+			reservedCpu := float64(rQuantity.Value()) - allocated.MilliCPU
+			if reservedCpu > 0 {
+				inqueue.MilliCPU = reservedCpu
+			}
+		case v1.ResourceMemory:
+			reservedMemory := float64(rQuantity.Value()) - allocated.Memory
+			if reservedMemory > 0 {
+				inqueue.Memory = reservedMemory
+			}
+		default:
+			if inqueue.ScalarResources == nil {
+				inqueue.ScalarResources = make(map[v1.ResourceName]float64)
+			}
+			if allocatedMount, ok := allocated.ScalarResources[rName]; !ok {
+				inqueue.ScalarResources[rName] = float64(rQuantity.Value())
+			} else {
+				reservedScalarRes := float64(rQuantity.Value()) - allocatedMount
+				if reservedScalarRes > 0 {
+					inqueue.ScalarResources[rName] = reservedScalarRes
+				}
+			}
+		}
+	}
+	return inqueue
+}