koordlet: cold memory fix usageWithHotPage and add usageWithPageCache (…

…#1699)

Signed-off-by: BUPT-wxq <[email protected]>

pkg/koordlet/metriccache/metric_resources.go

@@ -20,28 +20,32 @@ var (
 	defaultMetricFactory = NewMetricFactory()
 
 	// define all kinds of MetricResource
-	NodeCPUUsageMetric     = defaultMetricFactory.New(NodeMetricCPUUsage)
-	NodeMemoryUsageMetric  = defaultMetricFactory.New(NodeMetricMemoryUsage)
-	NodeGPUCoreUsageMetric = defaultMetricFactory.New(NodeMetricGPUCoreUsage).withPropertySchema(MetricPropertyGPUMinor, MetricPropertyGPUDeviceUUID)
-	NodeGPUMemUsageMetric  = defaultMetricFactory.New(NodeMetricGPUMemUsage).withPropertySchema(MetricPropertyGPUMinor, MetricPropertyGPUDeviceUUID)
-	NodeGPUMemTotalMetric  = defaultMetricFactory.New(NodeMetricGPUMemTotal).withPropertySchema(MetricPropertyGPUMinor, MetricPropertyGPUDeviceUUID)
+	NodeCPUUsageMetric                 = defaultMetricFactory.New(NodeMetricCPUUsage)
+	NodeMemoryUsageMetric              = defaultMetricFactory.New(NodeMetricMemoryUsage)
+	NodeMemoryUsageWithPageCacheMetric = defaultMetricFactory.New(NodeMemoryWithPageCacheUsage)
+	NodeGPUCoreUsageMetric             = defaultMetricFactory.New(NodeMetricGPUCoreUsage).withPropertySchema(MetricPropertyGPUMinor, MetricPropertyGPUDeviceUUID)
+	NodeGPUMemUsageMetric              = defaultMetricFactory.New(NodeMetricGPUMemUsage).withPropertySchema(MetricPropertyGPUMinor, MetricPropertyGPUDeviceUUID)
+	NodeGPUMemTotalMetric              = defaultMetricFactory.New(NodeMetricGPUMemTotal).withPropertySchema(MetricPropertyGPUMinor, MetricPropertyGPUDeviceUUID)
 
 	// define system resource usage as independent metric, although this can be calculate by node-sum(pod), but the time series are
 	// unaligned across different type of metric, which makes it hard to aggregate.
 	SystemCPUUsageMetric    = defaultMetricFactory.New(SysMetricCPUUsage)
 	SystemMemoryUsageMetric = defaultMetricFactory.New(SysMetricMemoryUsage)
 
-	PodCPUUsageMetric     = defaultMetricFactory.New(PodMetricCPUUsage).withPropertySchema(MetricPropertyPodUID)
-	PodMemUsageMetric     = defaultMetricFactory.New(PodMetricMemoryUsage).withPropertySchema(MetricPropertyPodUID)
+	PodCPUUsageMetric                 = defaultMetricFactory.New(PodMetricCPUUsage).withPropertySchema(MetricPropertyPodUID)
+	PodMemUsageMetric                 = defaultMetricFactory.New(PodMetricMemoryUsage).withPropertySchema(MetricPropertyPodUID)
+	PodMemoryUsageWithPageCacheMetric = defaultMetricFactory.New(PodMemoryWithPageCacheUsage).withPropertySchema(MetricPropertyPodUID)
+
 	PodCPUThrottledMetric = defaultMetricFactory.New(PodMetricCPUThrottled).withPropertySchema(MetricPropertyPodUID)
 	PodGPUCoreUsageMetric = defaultMetricFactory.New(PodMetricGPUCoreUsage).withPropertySchema(MetricPropertyPodUID, MetricPropertyGPUMinor, MetricPropertyGPUDeviceUUID)
 	PodGPUMemUsageMetric  = defaultMetricFactory.New(PodMetricGPUMemUsage).withPropertySchema(MetricPropertyPodUID, MetricPropertyGPUMinor, MetricPropertyGPUDeviceUUID)
 
-	ContainerCPUUsageMetric     = defaultMetricFactory.New(ContainerMetricCPUUsage).withPropertySchema(MetricPropertyContainerID)
-	ContainerMemUsageMetric     = defaultMetricFactory.New(ContainerMetricMemoryUsage).withPropertySchema(MetricPropertyContainerID)
-	ContainerGPUCoreUsageMetric = defaultMetricFactory.New(ContainerMetricGPUCoreUsage).withPropertySchema(MetricPropertyContainerID, MetricPropertyGPUMinor, MetricPropertyGPUDeviceUUID)
-	ContainerGPUMemUsageMetric  = defaultMetricFactory.New(ContainerMetricGPUMemUsage).withPropertySchema(MetricPropertyContainerID, MetricPropertyGPUMinor, MetricPropertyGPUDeviceUUID)
-	ContainerCPUThrottledMetric = defaultMetricFactory.New(ContainerMetricCPUThrottled).withPropertySchema(MetricPropertyContainerID)
+	ContainerCPUUsageMetric                 = defaultMetricFactory.New(ContainerMetricCPUUsage).withPropertySchema(MetricPropertyContainerID)
+	ContainerMemUsageMetric                 = defaultMetricFactory.New(ContainerMetricMemoryUsage).withPropertySchema(MetricPropertyContainerID)
+	ContainerMemoryUsageWithPageCacheMetric = defaultMetricFactory.New(ContainerMemoryWithPageCacheUsage).withPropertySchema(MetricPropertyContainerID)
+	ContainerGPUCoreUsageMetric             = defaultMetricFactory.New(ContainerMetricGPUCoreUsage).withPropertySchema(MetricPropertyContainerID, MetricPropertyGPUMinor, MetricPropertyGPUDeviceUUID)
+	ContainerGPUMemUsageMetric              = defaultMetricFactory.New(ContainerMetricGPUMemUsage).withPropertySchema(MetricPropertyContainerID, MetricPropertyGPUMinor, MetricPropertyGPUDeviceUUID)
+	ContainerCPUThrottledMetric             = defaultMetricFactory.New(ContainerMetricCPUThrottled).withPropertySchema(MetricPropertyContainerID)
 	// cold memory metrics
 	NodeMemoryWithHotPageUsageMetric      = defaultMetricFactory.New(NodeMemoryWithHotPageUsage)
 	PodMemoryWithHotPageUsageMetric       = defaultMetricFactory.New(PodMemoryWithHotPageUsage).withPropertySchema(MetricPropertyPodUID)

pkg/koordlet/metriccache/metric_types.go

@@ -36,11 +36,12 @@ const (
 type MetricKind string
 
 const (
-	NodeMetricCPUUsage     MetricKind = "node_cpu_usage"
-	NodeMetricMemoryUsage  MetricKind = "node_memory_usage"
-	NodeMetricGPUCoreUsage MetricKind = "node_gpu_core_usage"
-	NodeMetricGPUMemUsage  MetricKind = "node_gpu_memory_usage"
-	NodeMetricGPUMemTotal  MetricKind = "node_gpu_memory_total"
+	NodeMetricCPUUsage           MetricKind = "node_cpu_usage"
+	NodeMetricMemoryUsage        MetricKind = "node_memory_usage"
+	NodeMemoryWithPageCacheUsage MetricKind = "node_memory_usage_with_page_cache"
+	NodeMetricGPUCoreUsage       MetricKind = "node_gpu_core_usage"
+	NodeMetricGPUMemUsage        MetricKind = "node_gpu_memory_usage"
+	NodeMetricGPUMemTotal        MetricKind = "node_gpu_memory_total"
 
 	SysMetricCPUUsage    MetricKind = "sys_cpu_usage"
 	SysMetricMemoryUsage MetricKind = "sys_memory_usage"
@@ -52,16 +53,18 @@ const (
 	PriorityMetricCPURealLimit MetricKind = "priority_cpu_real_limit"
 	PriorityMetricCPURequest   MetricKind = "priority_cpu_request"
 
-	PodMetricCPUUsage     MetricKind = "pod_cpu_usage"
-	PodMetricMemoryUsage  MetricKind = "pod_memory_usage"
-	PodMetricGPUCoreUsage MetricKind = "pod_gpu_core_usage"
-	PodMetricGPUMemUsage  MetricKind = "pod_gpu_memory_usage"
+	PodMetricCPUUsage           MetricKind = "pod_cpu_usage"
+	PodMetricMemoryUsage        MetricKind = "pod_memory_usage"
+	PodMemoryWithPageCacheUsage MetricKind = "pod_memory_usage_with_page_cache"
+	PodMetricGPUCoreUsage       MetricKind = "pod_gpu_core_usage"
+	PodMetricGPUMemUsage        MetricKind = "pod_gpu_memory_usage"
 	// PodMetricGPUMemTotal       MetricKind = "pod_gpu_memory_total"
 
-	ContainerMetricCPUUsage     MetricKind = "container_cpu_usage"
-	ContainerMetricMemoryUsage  MetricKind = "container_memory_usage"
-	ContainerMetricGPUCoreUsage MetricKind = "container_gpu_core_usage"
-	ContainerMetricGPUMemUsage  MetricKind = "container_gpu_memory_usage"
+	ContainerMetricCPUUsage           MetricKind = "container_cpu_usage"
+	ContainerMetricMemoryUsage        MetricKind = "container_memory_usage"
+	ContainerMemoryWithPageCacheUsage MetricKind = "container_memory_usage_with_page_cache"
+	ContainerMetricGPUCoreUsage       MetricKind = "container_gpu_core_usage"
+	ContainerMetricGPUMemUsage        MetricKind = "container_gpu_memory_usage"
 	// ContainerMetricGPUMemTotal       MetricKind = "container_gpu_memory_total"
 
 	PodMetricCPUThrottled       MetricKind = "pod_cpu_throttled"

pkg/koordlet/metricsadvisor/collectors/coldmemoryresource/cold_page_collector.go

@@ -33,6 +33,7 @@ type nonColdPageCollector struct {
 func New(opt *framework.Options) framework.Collector {
 	// check whether support kidled cold page info collector
 	if system.IsKidledSupport() {
+		kidledConfig := system.NewDefaultKidledConfig()
 		return &kidledcoldPageCollector{
 			collectInterval: opt.Config.ColdPageCollectorInterval,
 			cgroupReader:    opt.CgroupReader,
@@ -42,6 +43,7 @@ func New(opt *framework.Options) framework.Collector {
 			appendableDB: opt.MetricCache,
 			metricDB:     opt.MetricCache,
 			started:      atomic.NewBool(false),
+			coldBoundary: kidledConfig.KidledColdBoundary,
 		}
 	}
 	// TODO(BUPT-wxq): check kstaled cold page collector

pkg/koordlet/metricsadvisor/collectors/coldmemoryresource/cold_page_collector_test.go

@@ -82,6 +82,7 @@ func Test_NewColdPageCollector(t *testing.T) {
 				appendableDB:    opt.MetricCache,
 				metricDB:        opt.MetricCache,
 				started:         atomic.NewBool(false),
+				coldBoundary:    3,
 			},
 		},
 	}

pkg/koordlet/metricsadvisor/collectors/coldmemoryresource/cold_page_kidled.go

@@ -43,6 +43,7 @@ type kidledcoldPageCollector struct {
 	podFilter       framework.PodFilter
 	appendableDB    metriccache.Appendable
 	metricDB        metriccache.MetricCache
+	coldBoundary    int
 }
 
 func (k *kidledcoldPageCollector) Run(stopCh <-chan struct{}) {
@@ -68,6 +69,7 @@ func (k *kidledcoldPageCollector) Enabled() bool {
 			return false
 		}
 		system.SetIsStartColdMemory(true)
+		system.SetKidledColdBoundary(k.coldBoundary)
 		return true
 	}
 	return false
@@ -121,7 +123,7 @@ func (k *kidledcoldPageCollector) collectNodeColdPageInfo() ([]metriccache.Metri
 	}
 	coldPageMetrics = append(coldPageMetrics, nodeColdPageMetrics)
 
-	memUsageWithHotPageBytes, err := koordletutil.GetNodeMemUsageWithHotPage(nodeColdPageBytes)
+	memUsageWithHotPageBytes, err := koordletutil.GetNodeMemUsageWithHotPageCache(nodeColdPageBytes)
 	if err != nil {
 		return nil, err
 	}
@@ -167,7 +169,7 @@ func (k *kidledcoldPageCollector) collectPodsColdPageInfo() ([]metriccache.Metri
 		}
 		coldMetrics = append(coldMetrics, podColdPageMetrics)
 
-		podMemUsageWithHotPageBytes, err := koordletutil.GetPodMemUsageWithHotPage(k.cgroupReader, podCgroupDir, podColdPageBytes)
+		podMemUsageWithHotPageBytes, err := koordletutil.GetPodMemUsageWithHotPageCache(k.cgroupReader, podCgroupDir, podColdPageBytes)
 		if err != nil {
 			klog.Warningf("failed to collect pod usage for Memory err: %s pod: %s/%s", err, pod.Namespace, pod.Name)
 			continue
@@ -220,7 +222,7 @@ func (k *kidledcoldPageCollector) collectContainersColdPageInfo(meta *statesinfo
 		}
 		coldMetrics = append(coldMetrics, containerColdPageMetrics)
 
-		containerMemUsageWithHotPageBytes, err := koordletutil.GetContainerMemUsageWithHotPage(k.cgroupReader, containerCgroupDir, containerColdPageBytes)
+		containerMemUsageWithHotPageBytes, err := koordletutil.GetContainerMemUsageWithHotPageCache(k.cgroupReader, containerCgroupDir, containerColdPageBytes)
 		if err != nil {
 			return nil, err
 		}

pkg/koordlet/metricsadvisor/collectors/coldmemoryresource/cold_page_kidled_test.go

@@ -485,6 +485,7 @@ func Test_collectNodeColdPageInfo(t *testing.T) {
 	helper.SetResourcesSupported(true, system.MemoryIdlePageStats)
 	helper.WriteCgroupFileContents("", system.MemoryIdlePageStats, idleInfoContentStr)
 	helper.WriteProcSubFileContents(system.ProcMemInfoName, meminfo)
+	kidledConfig := system.NewDefaultKidledConfig()
 	c := &kidledcoldPageCollector{
 		collectInterval: 5 * time.Second,
 		cgroupReader:    resourceexecutor.NewCgroupReader(),
@@ -493,6 +494,7 @@ func Test_collectNodeColdPageInfo(t *testing.T) {
 		appendableDB:    metricCache,
 		metricDB:        metricCache,
 		started:         atomic.NewBool(false),
+		coldBoundary:    kidledConfig.KidledColdBoundary,
 	}
 	testNow := time.Now()
 	metrics, err := c.collectNodeColdPageInfo()
@@ -509,8 +511,8 @@ func Test_collectNodeColdPageInfo(t *testing.T) {
 	}
 	assert.NoError(t, err)
 	got1, got2 := testGetNodeMetrics(t, c.metricDB, testNow, 5*time.Second)
-	assert.Equal(t, float64(18446744073419457000), got1)
-	assert.Equal(t, float64(1363836928), got2)
+	assert.Equal(t, float64(7.33569024e+08), got1)
+	assert.Equal(t, float64(3.401728e+08), got2)
 	// test collect failed
 	helper.WriteCgroupFileContents("", system.MemoryIdlePageStats, ``)
 	helper.WriteProcSubFileContents(system.ProcMemInfoName, ``)