scheduler: sort gang of same gangGroup by gangId

Signed-off-by: wangjianyu.wjy <wangjianyu.wjy@alibaba-inc.com>

pkg/scheduler/plugins/coscheduling/core/gang.go

@@ -390,12 +390,9 @@ func (gang *Gang) setScheduleCycleValid(valid bool) {
 
 	if !valid && !gang.ScheduleCycleValid {
 		/*
-				let's explain why by following example, there are three gang of one group: A, B and C, every gang group have min-member of 10 pod, noted as A1-A10, B1-B10, C1-C10.
-				1. A1-A5 assumed in gangA scheduleCycle 1, B1-B5 assumed in gangA scheduleCycle 1, C1-C5 assumed in gangA scheduleCycle 1, C6 filter failed due to insufficient resource, then A6-10 failed due to cycle invalid, C7-C10 failed due to cycle invalid, B6-B9 failed due to cycle invalid
-				2. A1-A5 assumed in gangA scheduleCycle 2, C1-C5 assumed in gangA scheduleCycle 2, C6 failed due to insufficient resource and gangA\B\C cycle set to invalid,
-				3. then A6-A10,C7-C10 failed due to cycle invalid
-				4. then B10 failed due to cycle invalid, B1 comes and find its gang scheduling cycle 2 can be valid, so it's assumed.
-				5. however all it's sibling will come until next round of all gangs, these gangs will face insufficient resource due to pre-allocated of B10
+			let's explain why by following example, there are three gang of one group: F, G, every gang group have min-member of 10 pod, noted as F1-F10, G1-G10.
+			1. F1 failed due to insufficient resource, F2-F10 failed due to cycle invalid,
+			2. then G1-G10 assumed, however all it's sibling will come until next round of all gangs, these gangs will face insufficient resource due to pre-allocated of G1-G10
 
 			TODO this logic can be optimized by give gangs of same group same scheduling cycle
 		*/

pkg/scheduler/plugins/coscheduling/coscheduling.go

@@ -140,20 +140,15 @@ func (cs *Coscheduling) Less(podInfo1, podInfo2 *framework.QueuedPodInfo) bool {
 	}
 	gangId1 := util.GetId(podInfo1.Pod.Namespace, util.GetGangNameByPod(podInfo1.Pod))
 	gangId2 := util.GetId(podInfo2.Pod.Namespace, util.GetGangNameByPod(podInfo2.Pod))
-	// for member gang of same gangGroup, the gang that haven’t been satisfied yet take precedence
+
 	if gangId1 != gangId2 {
-		isGang1Satisfied := cs.pgMgr.IsGangMinSatisfied(podInfo1.Pod)
-		isGang2Satisfied := cs.pgMgr.IsGangMinSatisfied(podInfo2.Pod)
-		if isGang1Satisfied != isGang2Satisfied {
-			return !isGang1Satisfied
-		}
-	} else {
-		// for member pod of same gang, the pod with the smaller scheduling cycle take precedence so that gang scheduling cycle can be valid and iterated
-		childScheduleCycle1 := cs.pgMgr.GetChildScheduleCycle(podInfo1.Pod)
-		childScheduleCycle2 := cs.pgMgr.GetChildScheduleCycle(podInfo2.Pod)
-		if childScheduleCycle1 != childScheduleCycle2 {
-			return childScheduleCycle1 < childScheduleCycle2
-		}
+		return gangId1 < gangId2
+	}
+	// for member pod of same gang, the pod with the smaller scheduling cycle take precedence so that gang scheduling cycle can be valid and iterated
+	childScheduleCycle1 := cs.pgMgr.GetChildScheduleCycle(podInfo1.Pod)
+	childScheduleCycle2 := cs.pgMgr.GetChildScheduleCycle(podInfo2.Pod)
+	if childScheduleCycle1 != childScheduleCycle2 {
+		return childScheduleCycle1 < childScheduleCycle2
 	}
 	return podInfo1.Timestamp.Before(podInfo2.Timestamp)
 }

pkg/scheduler/plugins/coscheduling/coscheduling_test.go

@@ -915,7 +915,7 @@ func TestFairness(t *testing.T) {
 		"gangJ": {"default/gangH", "default/gangI", "default/gangJ"},
 	}
 	gangMinRequiredNums := []int{20, 10, 32, 20, 20, 18, 43, 20, 30, 20}
-	gangInjectFilterError := []bool{false, false, true, false, false, true, true, false, false, true}
+	gangInjectFilterError := []bool{false, false, true, false, false, true, false, false, false, true}
 	var gangInjectFilterErrorIndex []int
 	for _, gangMinRequiredNum := range gangMinRequiredNums {
 		gangInjectFilterErrorIndex = append(gangInjectFilterErrorIndex, rand.Intn(gangMinRequiredNum))
@@ -996,8 +996,8 @@ func TestFairness(t *testing.T) {
 		ctx.Done(),
 		scheduler.WithProfiles(cfg.Profiles...),
 		scheduler.WithFrameworkOutOfTreeRegistry(registry),
-		scheduler.WithPodInitialBackoffSeconds(1),
-		scheduler.WithPodMaxBackoffSeconds(1),
+		scheduler.WithPodInitialBackoffSeconds(0),
+		scheduler.WithPodMaxBackoffSeconds(0),
 		scheduler.WithPodMaxInUnschedulablePodsDuration(0),
 	)
 	assert.NoError(t, err)