Merge pull request #18382 from yujuhong/resource_usage

cluster/gce/config-test.sh

@@ -71,7 +71,7 @@ ENABLE_CLUSTER_MONITORING="${KUBE_ENABLE_CLUSTER_MONITORING:-influxdb}"
 TEST_CLUSTER_LOG_LEVEL="${TEST_CLUSTER_LOG_LEVEL:---v=4}"
 TEST_CLUSTER_RESYNC_PERIOD="${TEST_CLUSTER_RESYNC_PERIOD:---min-resync-period=3m}"
 
-KUBELET_TEST_ARGS="--max-pods=100 $TEST_CLUSTER_LOG_LEVEL"
+KUBELET_TEST_ARGS="--max-pods=110 $TEST_CLUSTER_LOG_LEVEL"
 APISERVER_TEST_ARGS="--runtime-config=extensions/v1beta1 ${TEST_CLUSTER_LOG_LEVEL}"
 CONTROLLER_MANAGER_TEST_ARGS="${TEST_CLUSTER_LOG_LEVEL} ${TEST_CLUSTER_RESYNC_PERIOD}"
 SCHEDULER_TEST_ARGS="${TEST_CLUSTER_LOG_LEVEL}"

hack/jenkins/e2e.sh

@@ -289,7 +289,7 @@ GKE_REQUIRED_SKIP_TESTS=(
     "Nodes"
     "Etcd\sFailure"
     "MasterCerts"
-    "experimental\sresource\susage\stracking" # Expect --max-pods=100
+    "experimental\sresource\susage\stracking" # Expect --max-pods=110
     "ServiceLoadBalancer" # issue: #16602
     "Shell"
     # Alpha features, remove from skip when these move to beta
@@ -312,7 +312,7 @@ GKE_DEFAULT_SKIP_TESTS=(
 
 # Tests which cannot be run on AWS.
 AWS_REQUIRED_SKIP_TESTS=(
-    "experimental\sresource\susage\stracking" # Expect --max-pods=100
+    "experimental\sresource\susage\stracking" # Expect --max-pods=110
     "GCE\sL7\sLoadBalancer\sController" # GCE L7 loadbalancing
 )
 

test/e2e/kubelet_perf.go

@@ -18,6 +18,7 @@ package e2e
 
 import (
 	"fmt"
+	"strings"
 	"time"
 
 	"k8s.io/kubernetes/pkg/api/unversioned"
@@ -31,13 +32,18 @@ import (
 
 const (
 	// Interval to poll /stats/container on a node
-	containerStatsPollingPeriod = 10 * time.Second
+	containerStatsPollingPeriod = 3 * time.Second
 	// The monitoring time for one test.
-	monitoringTime = 30 * time.Minute
+	monitoringTime = 20 * time.Minute
 	// The periodic reporting period.
 	reportingPeriod = 5 * time.Minute
 )
 
+type resourceTest struct {
+	podsPerNode int
+	limits      containersCPUSummary
+}
+
 func logPodsOnNodes(c *client.Client, nodeNames []string) {
 	for _, n := range nodeNames {
 		podList, err := GetKubeletPods(c, n)
@@ -49,7 +55,7 @@ func logPodsOnNodes(c *client.Client, nodeNames []string) {
 	}
 }
 
-func runResourceTrackingTest(framework *Framework, podsPerNode int, nodeNames sets.String, resourceMonitor *resourceMonitor) {
+func runResourceTrackingTest(framework *Framework, podsPerNode int, nodeNames sets.String, rm *resourceMonitor, expected map[string]map[float64]float64) {
 	numNodes := nodeNames.Len()
 	totalPods := podsPerNode * numNodes
 	By(fmt.Sprintf("Creating a RC of %d pods and wait until all pods of this RC are running", totalPods))
@@ -65,8 +71,8 @@ func runResourceTrackingTest(framework *Framework, podsPerNode int, nodeNames se
 	})).NotTo(HaveOccurred())
 
 	// Log once and flush the stats.
-	resourceMonitor.LogLatest()
-	resourceMonitor.Reset()
+	rm.LogLatest()
+	rm.Reset()
 
 	By("Start monitoring resource usage")
 	// Periodically dump the cpu summary until the deadline is met.
@@ -76,8 +82,6 @@ func runResourceTrackingTest(framework *Framework, podsPerNode int, nodeNames se
 	// entries if we plan to monitor longer (e.g., 8 hours).
 	deadline := time.Now().Add(monitoringTime)
 	for time.Now().Before(deadline) {
-		Logf("Still running...%v left", deadline.Sub(time.Now()))
-		time.Sleep(reportingPeriod)
 		timeLeft := deadline.Sub(time.Now())
 		Logf("Still running...%v left", timeLeft)
 		if timeLeft < reportingPeriod {
@@ -90,17 +94,54 @@ func runResourceTrackingTest(framework *Framework, podsPerNode int, nodeNames se
 
 	By("Reporting overall resource usage")
 	logPodsOnNodes(framework.Client, nodeNames.List())
-	resourceMonitor.LogCPUSummary()
-	resourceMonitor.LogLatest()
+	rm.LogLatest()
+
+	summary := rm.GetCPUSummary()
+	Logf("%s", rm.FormatCPUSummary(summary))
+	verifyCPULimits(expected, summary)
 
 	By("Deleting the RC")
 	DeleteRC(framework.Client, framework.Namespace.Name, rcName)
 }
 
+func verifyCPULimits(expected containersCPUSummary, actual nodesCPUSummary) {
+	if expected == nil {
+		return
+	}
+	var errList []string
+	for nodeName, perNodeSummary := range actual {
+		var nodeErrs []string
+		for cName, expectedResult := range expected {
+			perContainerSummary, ok := perNodeSummary[cName]
+			if !ok {
+				nodeErrs = append(nodeErrs, fmt.Sprintf("container %q: missing", cName))
+				continue
+			}
+			for p, expectedValue := range expectedResult {
+				actualValue, ok := perContainerSummary[p]
+				if !ok {
+					nodeErrs = append(nodeErrs, fmt.Sprintf("container %q: missing percentile %v", cName, p))
+					continue
+				}
+				if actualValue > expectedValue {
+					nodeErrs = append(nodeErrs, fmt.Sprintf("container %q: expected %.0fth%% usage < %.3f; got %.3f",
+						cName, p*100, expectedValue, actualValue))
+				}
+			}
+		}
+		if len(nodeErrs) > 0 {
+			errList = append(errList, fmt.Sprintf("node %v:\n %s", nodeName, strings.Join(nodeErrs, ", ")))
+		}
+	}
+	if len(errList) > 0 {
+		Failf("CPU usage exceeding limits:\n %s", strings.Join(errList, "\n"))
+	}
+}
+
 var _ = Describe("Kubelet", func() {
 	var nodeNames sets.String
 	framework := NewFramework("kubelet-perf")
-	var resourceMonitor *resourceMonitor
+	var rm *resourceMonitor
 
 	BeforeEach(func() {
 		nodes, err := framework.Client.Nodes().List(unversioned.ListOptions{})
@@ -109,33 +150,46 @@ var _ = Describe("Kubelet", func() {
 		for _, node := range nodes.Items {
 			nodeNames.Insert(node.Name)
 		}
-		resourceMonitor = newResourceMonitor(framework.Client, targetContainers(), containerStatsPollingPeriod)
-		resourceMonitor.Start()
+		rm = newResourceMonitor(framework.Client, targetContainers(), containerStatsPollingPeriod)
+		rm.Start()
 	})
 
 	AfterEach(func() {
-		resourceMonitor.Stop()
+		rm.Stop()
 	})
-
 	Describe("regular resource usage tracking", func() {
-		density := []int{0, 35}
-		for i := range density {
-			podsPerNode := density[i]
+		rTests := []resourceTest{
+			{podsPerNode: 0,
+				limits: containersCPUSummary{
+					"/kubelet":       {0.50: 0.05, 0.95: 0.15},
+					"/docker-daemon": {0.50: 0.03, 0.95: 0.06},
+				},
+			},
+			{podsPerNode: 40,
+				limits: containersCPUSummary{
+					"/kubelet":       {0.50: 0.15, 0.95: 0.35},
+					"/docker-daemon": {0.50: 0.06, 0.95: 0.30},
+				},
+			},
+		}
+		for _, testArg := range rTests {
+			itArg := testArg
+			podsPerNode := itArg.podsPerNode
 			name := fmt.Sprintf(
-				"over %v with %d pods per node", monitoringTime, podsPerNode)
+				"for %d pods per node over %v", podsPerNode, monitoringTime)
 			It(name, func() {
-				runResourceTrackingTest(framework, podsPerNode, nodeNames, resourceMonitor)
+				runResourceTrackingTest(framework, podsPerNode, nodeNames, rm, itArg.limits)
 			})
 		}
 	})
 	Describe("experimental resource usage tracking", func() {
-		density := []int{50}
+		density := []int{100}
 		for i := range density {
 			podsPerNode := density[i]
 			name := fmt.Sprintf(
-				"over %v with %d pods per node.", monitoringTime, podsPerNode)
+				"for %d pods per node over %v", podsPerNode, monitoringTime)
 			It(name, func() {
-				runResourceTrackingTest(framework, podsPerNode, nodeNames, resourceMonitor)
+				runResourceTrackingTest(framework, podsPerNode, nodeNames, rm, nil)
 			})
 		}
 	})

test/e2e/kubelet_stats.go

@@ -692,7 +692,7 @@ func (r *resourceMonitor) Start() {
 	}
 	r.collectors = make(map[string]*resourceCollector, 0)
 	for _, node := range nodes.Items {
-		collector := newResourceCollector(r.client, node.Name, r.containers, pollInterval)
+		collector := newResourceCollector(r.client, node.Name, r.containers, r.pollingInterval)
 		r.collectors[node.Name] = collector
 		collector.Start()
 	}
@@ -716,33 +716,64 @@ func (r *resourceMonitor) LogLatest() {
 	}
 }
 
-func (r *resourceMonitor) LogCPUSummary() {
+// containersCPUSummary is indexed by the container name with each entry a
+// (percentile, value) map.
+type containersCPUSummary map[string]map[float64]float64
+
+// nodesCPUSummary is indexed by the node name with each entry a
+// containersCPUSummary map.
+type nodesCPUSummary map[string]containersCPUSummary
+
+func (r *resourceMonitor) FormatCPUSummary(summary nodesCPUSummary) string {
 	// Example output for a node (the percentiles may differ):
-	// CPU usage of containers on node "e2e-test-yjhong-minion-0vj7":
+	// CPU usage of containers on node "e2e-test-foo-minion-0vj7":
 	// container        5th%  50th% 90th% 95th%
 	// "/"              0.051 0.159 0.387 0.455
 	// "/docker-daemon" 0.000 0.000 0.146 0.166
 	// "/kubelet"       0.036 0.053 0.091 0.154
 	// "/system"        0.001 0.001 0.001 0.002
+	var summaryStrings []string
 	var header []string
 	header = append(header, "container")
 	for _, p := range percentiles {
 		header = append(header, fmt.Sprintf("%.0fth%%", p*100))
 	}
-	for nodeName, collector := range r.collectors {
+	for nodeName, containers := range summary {
 		buf := &bytes.Buffer{}
 		w := tabwriter.NewWriter(buf, 1, 0, 1, ' ', 0)
 		fmt.Fprintf(w, "%s\n", strings.Join(header, "\t"))
 		for _, containerName := range targetContainers() {
-			data := collector.GetBasicCPUStats(containerName)
 			var s []string
 			s = append(s, fmt.Sprintf("%q", containerName))
+			data, ok := containers[containerName]
 			for _, p := range percentiles {
-				s = append(s, fmt.Sprintf("%.3f", data[p]))
+				value := "N/A"
+				if ok {
+					value = fmt.Sprintf("%.3f", data[p])
+				}
+				s = append(s, value)
 			}
 			fmt.Fprintf(w, "%s\n", strings.Join(s, "\t"))
 		}
 		w.Flush()
-		Logf("\nCPU usage of containers on node %q:\n%s", nodeName, buf.String())
+		summaryStrings = append(summaryStrings, fmt.Sprintf("CPU usage of containers on node %q\n:%s", nodeName, buf.String()))
 	}
+	return strings.Join(summaryStrings, "\n")
+}
+
+func (r *resourceMonitor) LogCPUSummary() {
+	summary := r.GetCPUSummary()
+	Logf(r.FormatCPUSummary(summary))
+}
+
+func (r *resourceMonitor) GetCPUSummary() nodesCPUSummary {
+	result := make(nodesCPUSummary)
+	for nodeName, collector := range r.collectors {
+		result[nodeName] = make(containersCPUSummary)
+		for _, containerName := range targetContainers() {
+			data := collector.GetBasicCPUStats(containerName)
+			result[nodeName][containerName] = data
+		}
+	}
+	return result
 }