[YUNIKORN-2930] Add time consumption metrics for scheduling/tryNode cycle (#1017)

Add two new metrics that observe:
* each scheduling cycle, regardless of whether a pod was scheduled.
* total time taken to find a node for a given pod.

Add metrics to test

Closes: #1017

Signed-off-by: Wilfred Spiegelenburg <[email protected]>
(cherry picked from commit 29e95b6)

Author: kaichiachen

pkg/metrics/scheduler.go

@@ -62,9 +62,11 @@ type SchedulerMetrics struct {
 	node                  *prometheus.GaugeVec
 	nodeResourceUsage     map[string]*prometheus.GaugeVec
 	schedulingLatency     prometheus.Histogram
+	schedulingCycle       prometheus.Histogram
 	sortingLatency        *prometheus.HistogramVec
 	tryNodeLatency        prometheus.Histogram
 	tryPreemptionLatency  prometheus.Histogram
+	tryNodeEvaluation     prometheus.Histogram
 	lock                  locking.RWMutex
 }
 
@@ -117,6 +119,17 @@ func InitSchedulerMetrics() *SchedulerMetrics {
 			Buckets:   prometheus.ExponentialBuckets(0.0001, 10, 8), // start from 0.1ms
 		},
 	)
+
+	s.schedulingCycle = prometheus.NewHistogram(
+		prometheus.HistogramOpts{
+			Namespace: Namespace,
+			Subsystem: SchedulerSubsystem,
+			Name:      "scheduling_cycle_milliseconds",
+			Help:      "Time taken for a scheduling cycle, in seconds.",
+			Buckets:   prometheus.ExponentialBuckets(0.0001, 10, 8),
+		},
+	)
+
 	s.sortingLatency = prometheus.NewHistogramVec(
 		prometheus.HistogramOpts{
 			Namespace: Namespace,
@@ -136,6 +149,16 @@ func InitSchedulerMetrics() *SchedulerMetrics {
 		},
 	)
 
+	s.tryNodeEvaluation = prometheus.NewHistogram(
+		prometheus.HistogramOpts{
+			Namespace: Namespace,
+			Subsystem: SchedulerSubsystem,
+			Name:      "trynode_evaluation_milliseconds",
+			Help:      "Time taken to evaluate nodes for a pod, in seconds.",
+			Buckets:   prometheus.ExponentialBuckets(0.0001, 10, 8),
+		},
+	)
+
 	s.tryPreemptionLatency = prometheus.NewHistogram(
 		prometheus.HistogramOpts{
 			Namespace: Namespace,
@@ -155,6 +178,8 @@ func InitSchedulerMetrics() *SchedulerMetrics {
 		s.schedulingLatency,
 		s.sortingLatency,
 		s.tryNodeLatency,
+		s.schedulingCycle,
+		s.tryNodeEvaluation,
 		s.tryPreemptionLatency,
 	}
 	for _, metric := range metricsList {
@@ -182,6 +207,10 @@ func (m *SchedulerMetrics) ObserveSchedulingLatency(start time.Time) {
 	m.schedulingLatency.Observe(SinceInSeconds(start))
 }
 
+func (m *SchedulerMetrics) ObserveSchedulingCycle(start time.Time) {
+	m.schedulingCycle.Observe(SinceInSeconds(start))
+}
+
 func (m *SchedulerMetrics) ObserveAppSortingLatency(start time.Time) {
 	m.sortingLatency.WithLabelValues(SortingApp).Observe(SinceInSeconds(start))
 }
@@ -194,6 +223,10 @@ func (m *SchedulerMetrics) ObserveTryNodeLatency(start time.Time) {
 	m.tryNodeLatency.Observe(SinceInSeconds(start))
 }
 
+func (m *SchedulerMetrics) ObserveTryNodeEvaluation(start time.Time) {
+	m.tryNodeEvaluation.Observe(SinceInSeconds(start))
+}
+
 func (m *SchedulerMetrics) ObserveTryPreemptionLatency(start time.Time) {
 	m.tryPreemptionLatency.Observe(SinceInSeconds(start))
 }

pkg/metrics/scheduler_test.go

@@ -168,6 +168,22 @@ func TestSchedulerApplicationsFailed(t *testing.T) {
 	verifyMetric(t, 1, "failed", "yunikorn_scheduler_application_total", dto.MetricType_GAUGE, "state")
 }
 
+func TestSchedulingCycle(t *testing.T) {
+	sm = getSchedulerMetrics(t)
+	defer unregisterMetrics()
+
+	sm.ObserveSchedulingCycle(time.Now().Add(-1 * time.Minute))
+	verifyHistogram(t, "scheduling_cycle_milliseconds", 60, 1)
+}
+
+func TestTryNodeEvaluation(t *testing.T) {
+	sm = getSchedulerMetrics(t)
+	defer unregisterMetrics()
+
+	sm.ObserveTryNodeEvaluation(time.Now().Add(-1 * time.Minute))
+	verifyHistogram(t, "trynode_evaluation_milliseconds", 60, 1)
+}
+
 func getSchedulerMetrics(t *testing.T) *SchedulerMetrics {
 	unregisterMetrics()
 	return InitSchedulerMetrics()
@@ -223,7 +239,9 @@ func unregisterMetrics() {
 	prometheus.Unregister(sm.application)
 	prometheus.Unregister(sm.node)
 	prometheus.Unregister(sm.schedulingLatency)
+	prometheus.Unregister(sm.schedulingCycle)
 	prometheus.Unregister(sm.sortingLatency)
 	prometheus.Unregister(sm.tryNodeLatency)
+	prometheus.Unregister(sm.tryNodeEvaluation)
 	prometheus.Unregister(sm.tryPreemptionLatency)
 }

pkg/scheduler/context.go

@@ -120,6 +120,7 @@ func (cc *ClusterContext) setEventHandler(rmHandler handler.EventHandler) {
 func (cc *ClusterContext) schedule() bool {
 	// schedule each partition defined in the cluster
 	activity := false
+	scheduleCycleStart := time.Now()
 	for _, psc := range cc.GetPartitionMapClone() {
 		// if there are no resources in the partition just skip
 		if psc.root.GetMaxResource() == nil {
@@ -151,6 +152,7 @@ func (cc *ClusterContext) schedule() bool {
 			activity = true
 		}
 	}
+	metrics.GetSchedulerMetrics().ObserveSchedulingCycle(scheduleCycleStart)
 	return activity
 }
 

pkg/scheduler/objects/application.go

@@ -1446,6 +1446,7 @@ func (sa *Application) tryNodes(ask *Allocation, iterator NodeIterator) *Allocat
 	reserved := sa.reservations[allocKey]
 	var allocResult *AllocationResult
 	var predicateErrors map[string]int
+	tryNodeCycleStart := time.Now()
 	iterator.ForEachNode(func(node *Node) bool {
 		// skip the node if the node is not schedulable
 		if !node.IsSchedulable() {
@@ -1510,6 +1511,7 @@ func (sa *Application) tryNodes(ask *Allocation, iterator NodeIterator) *Allocat
 		}
 		return true
 	})
+	metrics.GetSchedulerMetrics().ObserveTryNodeEvaluation(tryNodeCycleStart)
 
 	if allocResult != nil {
 		return allocResult