Merge pull request #513 from karan/cpu_activity_metrics

add metrics for process stats

Author: k8s-ci-robot

config/system-stats-monitor.json

@@ -15,7 +15,19 @@
       },
       "cpu/load_15m": {
 				"displayName": "cpu/load_15m"
-			}
+      },
+      "system/processes_total": {
+        "displayName": "system/processes_total"
+      },
+      "system/procs_running": {
+        "displayName": "system/procs_running"
+      },
+      "system/procs_blocked": {
+        "displayName": "system/procs_blocked"
+      },
+      "system/interrupts_total": {
+        "displayName": "system/interrupts_total"
+      }
 		}
 	},
 	"disk": {

pkg/exporters/stackdriver/stackdriver_exporter.go

@@ -69,6 +69,10 @@ var NPDMetricToSDMetric = map[metrics.MetricID]string{
 	metrics.ProblemCounterID:        "compute.googleapis.com/guest/system/problem_count",
 	metrics.ProblemGaugeID:          "compute.googleapis.com/guest/system/problem_state",
 	metrics.OSFeatureID:             "compute.googleapis.com/guest/system/os_feature_enabled",
+	metrics.SystemProcessesTotal:    "kubernetes.io/internal/node/guest/system/processes_total",
+	metrics.SystemProcsRunning:      "kubernetes.io/internal/node/guest/system/procs_running",
+	metrics.SystemProcsBlocked:      "kubernetes.io/internal/node/guest/system/procs_blocked",
+	metrics.SystemInterruptsTotal:   "kubernetes.io/internal/node/guest/system/interrupts_total",
 }
 
 func getMetricTypeConversionFunction(customMetricPrefix string) func(*view.View) string {

pkg/systemstatsmonitor/README.md

@@ -28,6 +28,10 @@ Below metrics are collected from `cpu` component:
 * `cpu_load_1m`: CPU load average over the last 1 minute. Collected from [`/proc/loadavg`][/proc doc].
 * `cpu_load_5m`: CPU load average over the last 5 minutes. Collected from [`/proc/loadavg`][/proc doc].
 * `cpu_load_15m`: CPU load average over the last 15 minutes. Collected from [`/proc/loadavg`][/proc doc].
+* `system/processes_total`: Number of forks since boot.
+* `system/procs_running`: Number of processes currently running.
+* `system/procs_blocked`: Number of processes currently blocked.
+* `system/interrupts_total`: Total number of interrupts serviced (cumulative).
 
 [/proc doc]: http://man7.org/linux/man-pages/man5/proc.5.html
 

pkg/systemstatsmonitor/cpu_collector.go

@@ -18,10 +18,13 @@ package systemstatsmonitor
 
 import (
 	"github.com/golang/glog"
+	"github.com/prometheus/procfs"
 	"github.com/shirou/gopsutil/cpu"
+	"github.com/shirou/gopsutil/host"
 	"github.com/shirou/gopsutil/load"
 
 	ssmtypes "k8s.io/node-problem-detector/pkg/systemstatsmonitor/types"
+	"k8s.io/node-problem-detector/pkg/util"
 	"k8s.io/node-problem-detector/pkg/util/metrics"
 )
 
@@ -36,21 +39,37 @@ import (
 const clockTick float64 = 100.0
 
 type cpuCollector struct {
-	mRunnableTaskCount *metrics.Float64Metric
-	mUsageTime         *metrics.Float64Metric
-	mCpuLoad1m         *metrics.Float64Metric
-	mCpuLoad5m         *metrics.Float64Metric
-	mCpuLoad15m        *metrics.Float64Metric
+	tags map[string]string
+
+	mRunnableTaskCount     *metrics.Float64Metric
+	mUsageTime             *metrics.Float64Metric
+	mCpuLoad1m             *metrics.Float64Metric
+	mCpuLoad5m             *metrics.Float64Metric
+	mCpuLoad15m            *metrics.Float64Metric
+	mSystemProcessesTotal  *metrics.Int64Metric
+	mSystemProcsRunning    *metrics.Int64Metric
+	mSystemProcsBlocked    *metrics.Int64Metric
+	mSystemInterruptsTotal *metrics.Int64Metric
 
 	config *ssmtypes.CPUStatsConfig
 
 	lastUsageTime map[string]float64
 }
 
 func NewCPUCollectorOrDie(cpuConfig *ssmtypes.CPUStatsConfig) *cpuCollector {
-	cc := cpuCollector{config: cpuConfig}
+	cc := cpuCollector{tags: map[string]string{}, config: cpuConfig}
 
-	var err error
+	kernelVersion, err := host.KernelVersion()
+	if err != nil {
+		glog.Fatalf("Failed to retrieve kernel version: %v", err)
+	}
+	cc.tags["kernel_version"] = kernelVersion
+
+	osVersion, err := util.GetOSVersion()
+	if err != nil {
+		glog.Fatalf("Failed to retrieve OS version: %v", err)
+	}
+	cc.tags["os_version"] = osVersion
 
 	cc.mRunnableTaskCount, err = metrics.NewFloat64Metric(
 		metrics.CPURunnableTaskCountID,
@@ -107,6 +126,50 @@ func NewCPUCollectorOrDie(cpuConfig *ssmtypes.CPUStatsConfig) *cpuCollector {
 		glog.Fatalf("Error initializing metric for %q: %v", metrics.CPULoad15m, err)
 	}
 
+	cc.mSystemProcessesTotal, err = metrics.NewInt64Metric(
+		metrics.SystemProcessesTotal,
+		cpuConfig.MetricsConfigs[string(metrics.SystemProcessesTotal)].DisplayName,
+		"Number of forks since boot.",
+		"1",
+		metrics.Sum,
+		[]string{osVersionLabel, kernelVersionLabel})
+	if err != nil {
+		glog.Fatalf("Error initializing metric for %q: %v", metrics.SystemProcessesTotal, err)
+	}
+
+	cc.mSystemProcsRunning, err = metrics.NewInt64Metric(
+		metrics.SystemProcsRunning,
+		cpuConfig.MetricsConfigs[string(metrics.SystemProcsRunning)].DisplayName,
+		"Number of processes currently running.",
+		"1",
+		metrics.LastValue,
+		[]string{osVersionLabel, kernelVersionLabel})
+	if err != nil {
+		glog.Fatalf("Error initializing metric for %q: %v", metrics.SystemProcsRunning, err)
+	}
+
+	cc.mSystemProcsBlocked, err = metrics.NewInt64Metric(
+		metrics.SystemProcsBlocked,
+		cpuConfig.MetricsConfigs[string(metrics.SystemProcsBlocked)].DisplayName,
+		"Number of processes currently blocked.",
+		"1",
+		metrics.LastValue,
+		[]string{osVersionLabel, kernelVersionLabel})
+	if err != nil {
+		glog.Fatalf("Error initializing metric for %q: %v", metrics.SystemProcsBlocked, err)
+	}
+
+	cc.mSystemInterruptsTotal, err = metrics.NewInt64Metric(
+		metrics.SystemInterruptsTotal,
+		cpuConfig.MetricsConfigs[string(metrics.SystemInterruptsTotal)].DisplayName,
+		"Total number of interrupts serviced (cumulative).",
+		"1",
+		metrics.Sum,
+		[]string{osVersionLabel, kernelVersionLabel})
+	if err != nil {
+		glog.Fatalf("Error initializing metric for %q: %v", metrics.SystemInterruptsTotal, err)
+	}
+
 	cc.lastUsageTime = make(map[string]float64)
 
 	return &cc
@@ -174,11 +237,39 @@ func (cc *cpuCollector) recordUsage() {
 	cc.lastUsageTime["guest_nice"] = clockTick * timersStat.GuestNice
 }
 
+func (cc *cpuCollector) recordSystemStats() {
+	if cc.mSystemProcessesTotal == nil {
+		return
+	}
+	if cc.mSystemProcsRunning == nil {
+		return
+	}
+	if cc.mSystemProcsBlocked == nil {
+		return
+	}
+	if cc.mSystemInterruptsTotal == nil {
+		return
+	}
+
+	fs, err := procfs.NewFS("/proc")
+	stats, err := fs.Stat()
+	if err != nil {
+		glog.Errorf("Failed to retrieve cpu/process stats: %v", err)
+		return
+	}
+
+	cc.mSystemProcessesTotal.Record(cc.tags, int64(stats.ProcessCreated))
+	cc.mSystemProcsRunning.Record(cc.tags, int64(stats.ProcessesRunning))
+	cc.mSystemProcsBlocked.Record(cc.tags, int64(stats.ProcessesBlocked))
+	cc.mSystemInterruptsTotal.Record(cc.tags, int64(stats.IRQTotal))
+}
+
 func (cc *cpuCollector) collect() {
 	if cc == nil {
 		return
 	}
 
 	cc.recordLoad()
 	cc.recordUsage()
+	cc.recordSystemStats()
 }

pkg/systemstatsmonitor/labels.go

@@ -36,3 +36,9 @@ const featureLabel = "os_feature"
 
 // valueLabel labels the value for the features of the guest os system if required
 const valueLabel = "value"
+
+// osVersionLabel labels the OS
+const osVersionLabel = "os_version"
+
+// osVersionLabel labels the kernel version
+const kernelVersionLabel = "kernel_version"

pkg/systemstatsmonitor/osfeature_collector.go

@@ -142,6 +142,9 @@ func (ofc *osFeatureCollector) recordFeaturesFromModules(modules []system.Module
 }
 
 func (ofc *osFeatureCollector) collect() {
+	if ofc == nil || ofc.osFeature == nil {
+		return
+	}
 	cmdlineArgs, err := system.CmdlineArgs()
 	if err != nil {
 		glog.Fatalf("Error retrieving cmdline args: %v", err)

pkg/util/metrics/metric.go

@@ -42,6 +42,10 @@ const (
 	MemoryUnevictableUsedID MetricID = "memory/unevictable_used"
 	MemoryDirtyUsedID       MetricID = "memory/dirty_used"
 	OSFeatureID             MetricID = "system/os_feature"
+	SystemProcessesTotal    MetricID = "system/processes_total"
+	SystemProcsRunning      MetricID = "system/procs_running"
+	SystemProcsBlocked      MetricID = "system/procs_blocked"
+	SystemInterruptsTotal   MetricID = "system/interrupts_total"
 )
 
 var MetricMap MetricMapping