e2e: checking RSS memory for daemons

test/e2e/kubelet_perf.go

@@ -95,15 +95,53 @@ func runResourceTrackingTest(framework *Framework, podsPerNode int, nodeNames se
 	By("Reporting overall resource usage")
 	logPodsOnNodes(framework.Client, nodeNames.List())
 	rm.LogLatest()
+	usageSummary, err := rm.GetLatest()
+	Expect(err).NotTo(HaveOccurred())
+	Logf("%s", rm.FormatResourceUsage(usageSummary))
+	// TODO(yujuhong): Set realistic values after gathering enough data.
+	verifyMemoryLimits(resourceUsagePerContainer{
+		"/kubelet":       &containerResourceUsage{MemoryRSSInBytes: 500 * 1024 * 1024},
+		"/docker-daemon": &containerResourceUsage{MemoryRSSInBytes: 500 * 1024 * 1024},
+	}, usageSummary)
 
-	summary := rm.GetCPUSummary()
+	cpuSummary := rm.GetCPUSummary()
-	Logf("%s", rm.FormatCPUSummary(summary))
+	Logf("%s", rm.FormatCPUSummary(cpuSummary))
-	verifyCPULimits(expected, summary)
+	verifyCPULimits(expected, cpuSummary)
 
 	By("Deleting the RC")
 	DeleteRC(framework.Client, framework.Namespace.Name, rcName)
 }
 
+func verifyMemoryLimits(expected resourceUsagePerContainer, actual resourceUsagePerNode) {
+	if expected == nil {
+		return
+	}
+	var errList []string
+	for nodeName, nodeSummary := range actual {
+		var nodeErrs []string
+		for cName, expectedResult := range expected {
+			container, ok := nodeSummary[cName]
+			if !ok {
+				nodeErrs = append(nodeErrs, fmt.Sprintf("container %q: missing", cName))
+				continue
+			}
+
+			expectedValue := expectedResult.MemoryRSSInBytes
+			actualValue := container.MemoryRSSInBytes
+			if expectedValue != 0 && actualValue > expectedValue {
+				nodeErrs = append(nodeErrs, fmt.Sprintf("container %q: expected RSS memory (MB) < %d; got %d",
+					cName, 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"))
+	}
+}
+
 func verifyCPULimits(expected containersCPUSummary, actual nodesCPUSummary) {
 	if expected == nil {
 		return

test/e2e/kubelet_stats.go

@@ -36,6 +36,7 @@ import (
 	"k8s.io/kubernetes/pkg/kubelet/metrics"
 	"k8s.io/kubernetes/pkg/kubelet/server/stats"
 	"k8s.io/kubernetes/pkg/master/ports"
+	utilerrors "k8s.io/kubernetes/pkg/util/errors"
 	"k8s.io/kubernetes/pkg/util/sets"
 	"k8s.io/kubernetes/pkg/util/wait"
 )
@@ -198,6 +199,7 @@ type containerResourceUsage struct {
 	CPUUsageInCores         float64
 	MemoryUsageInBytes      int64
 	MemoryWorkingSetInBytes int64
+	MemoryRSSInBytes        uint64
 	// The interval used to calculate CPUUsageInCores.
 	CPUInterval time.Duration
 }
@@ -207,6 +209,7 @@ func (r *containerResourceUsage) isStrictlyGreaterThan(rhs *containerResourceUsa
 }
 
 type resourceUsagePerContainer map[string]*containerResourceUsage
+type resourceUsagePerNode map[string]resourceUsagePerContainer
 
 // getOneTimeResourceUsageOnNode queries the node's /stats/container endpoint
 // and returns the resource usage of all containerNames for the past
@@ -292,9 +295,9 @@ func formatResourceUsageStats(nodeName string, containerStats resourceUsagePerCo
 	// "/system"        0.007       119.88
 	buf := &bytes.Buffer{}
 	w := tabwriter.NewWriter(buf, 1, 0, 1, ' ', 0)
-	fmt.Fprintf(w, "container\tcpu(cores)\tmemory(MB)\n")
+	fmt.Fprintf(w, "container\tcpu(cores)\tmemory_working_set(MB)\tmemory_rss(MB)\n")
 	for name, s := range containerStats {
-		fmt.Fprintf(w, "%q\t%.3f\t%.2f\n", name, s.CPUUsageInCores, float64(s.MemoryWorkingSetInBytes)/(1024*1024))
+		fmt.Fprintf(w, "%q\t%.3f\t%.2f\t%.2f\n", name, s.CPUUsageInCores, float64(s.MemoryWorkingSetInBytes)/(1024*1024), float64(s.MemoryRSSInBytes)/(1024*1024))
 	}
 	w.Flush()
 	return fmt.Sprintf("Resource usage on node %q:\n%s", nodeName, buf.String())
@@ -364,6 +367,7 @@ func computeContainerResourceUsage(name string, oldStats, newStats *cadvisorapi.
 		CPUUsageInCores:         float64(newStats.Cpu.Usage.Total-oldStats.Cpu.Usage.Total) / float64(newStats.Timestamp.Sub(oldStats.Timestamp).Nanoseconds()),
 		MemoryUsageInBytes:      int64(newStats.Memory.Usage),
 		MemoryWorkingSetInBytes: int64(newStats.Memory.WorkingSet),
+		MemoryRSSInBytes:        newStats.Memory.RSS,
 		CPUInterval:             newStats.Timestamp.Sub(oldStats.Timestamp),
 	}
 }
@@ -437,20 +441,18 @@ func (r *resourceCollector) collectStats(oldStats map[string]*cadvisorapi.Contai
 	}
 }
 
-// LogLatest logs the latest resource usage of each container.
+func (r *resourceCollector) GetLatest() (resourceUsagePerContainer, error) {
-func (r *resourceCollector) LogLatest() {
 	r.lock.RLock()
 	defer r.lock.RUnlock()
-	stats := make(map[string]*containerResourceUsage)
+	stats := make(resourceUsagePerContainer)
 	for _, name := range r.containers {
 		contStats, ok := r.buffers[name]
 		if !ok || len(contStats) == 0 {
-			Logf("Resource usage on node %q is not ready yet", r.node)
+			return nil, fmt.Errorf("Resource usage on node %q is not ready yet", r.node)
-			return
 		}
 		stats[name] = contStats[len(contStats)-1]
 	}
-	Logf("\n%s", formatResourceUsageStats(r.node, stats))
+	return stats, nil
 }
 
 // Reset frees the stats and start over.
@@ -534,9 +536,33 @@ func (r *resourceMonitor) Reset() {
 }
 
 func (r *resourceMonitor) LogLatest() {
-	for _, collector := range r.collectors {
+	summary, err := r.GetLatest()
-		collector.LogLatest()
+	if err != nil {
+		Logf("%v", err)
 	}
+	r.FormatResourceUsage(summary)
+}
+
+func (r *resourceMonitor) FormatResourceUsage(s resourceUsagePerNode) string {
+	summary := []string{}
+	for node, usage := range s {
+		summary = append(summary, formatResourceUsageStats(node, usage))
+	}
+	return strings.Join(summary, "\n")
+}
+
+func (r *resourceMonitor) GetLatest() (resourceUsagePerNode, error) {
+	result := make(resourceUsagePerNode)
+	errs := []error{}
+	for key, collector := range r.collectors {
+		s, err := collector.GetLatest()
+		if err != nil {
+			errs = append(errs, err)
+			continue
+		}
+		result[key] = s
+	}
+	return result, utilerrors.NewAggregate(errs)
 }
 
 // containersCPUSummary is indexed by the container name with each entry a