Merge pull request #24542 from Random-Liu/versioned-perf-data

Automatic merge from submit-queue

Use mCPU as CPU usage unit, add version in PerfData, and fix memory usage bug.

Partially addressed #24436.

This PR:
1) Change the CPU usage unit to "mCPU"
2) Add version in PerfData, and perfdash will only support the newest version now.
3) Fix stupid mistake when calculating the memory usage average.

/cc @vishh

test/e2e/framework/kubelet_stats.go

@@ -669,7 +669,7 @@ func (r *ResourceMonitor) GetMasterNodeLatest(usagePerNode ResourceUsagePerNode)
 				nodeAvgUsage[c] = &ContainerResourceUsage{Name: usage.Name}
 			}
 			nodeAvgUsage[c].CPUUsageInCores += usage.CPUUsageInCores
-			nodeAvgUsage[c].MemoryRSSInBytes += usage.MemoryRSSInBytes
+			nodeAvgUsage[c].MemoryUsageInBytes += usage.MemoryUsageInBytes
 			nodeAvgUsage[c].MemoryWorkingSetInBytes += usage.MemoryWorkingSetInBytes
 			nodeAvgUsage[c].MemoryRSSInBytes += usage.MemoryRSSInBytes
 		}

test/e2e/framework/perf_util.go

@@ -25,9 +25,13 @@ import (
 // TODO(random-liu): Change the tests to actually use PerfData from the beginning instead of
 // translating one to the other here.
 
+// currentApiCallMetricsVersion is the current apicall performance metrics version. We should
+// bump up the version each time we make incompatible change to the metrics.
+const currentApiCallMetricsVersion = "v1"
+
 // ApiCallToPerfData transforms APIResponsiveness to PerfData.
 func ApiCallToPerfData(apicalls APIResponsiveness) *perftype.PerfData {
-	perfData := &perftype.PerfData{}
+	perfData := &perftype.PerfData{Version: currentApiCallMetricsVersion}
 	for _, apicall := range apicalls.APICalls {
 		item := perftype.DataItem{
 			Data: map[string]float64{
@@ -46,6 +50,10 @@ func ApiCallToPerfData(apicalls APIResponsiveness) *perftype.PerfData {
 	return perfData
 }
 
+// currentKubeletPerfMetricsVersion is the current kubelet performance metrics version. We should
+// bump up the version each time we make incompatible change to the metrics.
+const currentKubeletPerfMetricsVersion = "v1"
+
 // ResourceUsageToPerfData transforms ResourceUsagePerNode to PerfData. Notice that this function
 // only cares about memory usage, because cpu usage information will be extracted from NodesCPUSummary.
 func ResourceUsageToPerfData(usagePerNode ResourceUsagePerNode) *perftype.PerfData {
@@ -54,7 +62,7 @@ func ResourceUsageToPerfData(usagePerNode ResourceUsagePerNode) *perftype.PerfDa
 		for c, usage := range usages {
 			item := perftype.DataItem{
 				Data: map[string]float64{
-					"memory":     float64(usage.MemoryRSSInBytes) / (1024 * 1024),
+					"memory":     float64(usage.MemoryUsageInBytes) / (1024 * 1024),
 					"workingset": float64(usage.MemoryWorkingSetInBytes) / (1024 * 1024),
 					"rss":        float64(usage.MemoryRSSInBytes) / (1024 * 1024),
 				},
@@ -68,7 +76,10 @@ func ResourceUsageToPerfData(usagePerNode ResourceUsagePerNode) *perftype.PerfDa
 			items = append(items, item)
 		}
 	}
-	return &perftype.PerfData{DataItems: items}
+	return &perftype.PerfData{
+		Version:   currentKubeletPerfMetricsVersion,
+		DataItems: items,
+	}
 }
 
 // CPUUsageToPerfData transforms NodesCPUSummary to PerfData.
@@ -78,11 +89,11 @@ func CPUUsageToPerfData(usagePerNode NodesCPUSummary) *perftype.PerfData {
 		for c, usage := range usages {
 			data := map[string]float64{}
 			for perc, value := range usage {
-				data[fmt.Sprintf("Perc%02.0f", perc*100)] = value * 100
+				data[fmt.Sprintf("Perc%02.0f", perc*100)] = value * 1000
 			}
 			item := perftype.DataItem{
 				Data: data,
-				Unit: "%",
+				Unit: "mCPU",
 				Labels: map[string]string{
 					"node":      node,
 					"container": c,
@@ -92,13 +103,17 @@ func CPUUsageToPerfData(usagePerNode NodesCPUSummary) *perftype.PerfData {
 			items = append(items, item)
 		}
 	}
-	return &perftype.PerfData{DataItems: items}
+	return &perftype.PerfData{
+		Version:   currentKubeletPerfMetricsVersion,
+		DataItems: items,
+	}
 }
 
 // PrintPerfData prints the perfdata in json format with PerfResultTag prefix.
 // If an error occurs, nothing will be printed.
 func PrintPerfData(p *perftype.PerfData) {
+	// Notice that we must make sure the perftype.PerfResultEnd is in a new line.
 	if str := PrettyPrintJSON(p); str != "" {
-		Logf("%s", perftype.PerfResultTag+" "+str)
+		Logf("%s %s\n%s", perftype.PerfResultTag, str, perftype.PerfResultEnd)
 	}
 }

test/e2e/perftype/perftype.go

@@ -36,9 +36,16 @@ type DataItem struct {
 
 // PerfData contains all data items generated in current test.
 type PerfData struct {
+	// Version is the version of the metrics. The metrics consumer could use the version
+	// to detect metrics version change and decide what version to support.
+	Version   string     `json:"version"`
 	DataItems []DataItem `json:"dataItems"`
 }
 
 // PerfResultTag is the prefix of generated perfdata. Analyzing tools can find the perf result
 // with this tag.
 const PerfResultTag = "[Result:Performance]"
+
+// PerfResultEnd is the end of generated perfdata. Analyzing tools can find the end of the perf
+// result with this tag.
+const PerfResultEnd = "[Finish:Performance]"