Merge pull request #976 from sunya-ch/model-server

move model initialization after BPF attach
sustainable-computing-io · Oct 10, 2023 · 34f6562 · 34f6562
2 parents 8529097 + c9f69ea
commit 34f6562
Show file tree

Hide file tree

Showing 4 changed files with 25 additions and 11 deletions.
diff --git a/cmd/exporter/exporter.go b/cmd/exporter/exporter.go
@@ -32,7 +32,6 @@ import (
 	collector_metric "github.com/sustainable-computing-io/kepler/pkg/collector/metric"
 	"github.com/sustainable-computing-io/kepler/pkg/config"
 	"github.com/sustainable-computing-io/kepler/pkg/manager"
-	"github.com/sustainable-computing-io/kepler/pkg/model"
 	"github.com/sustainable-computing-io/kepler/pkg/power/accelerator/gpu"
 	"github.com/sustainable-computing-io/kepler/pkg/power/accelerator/qat"
 	"github.com/sustainable-computing-io/kepler/pkg/power/components"
@@ -229,14 +228,6 @@ func main() {
 		}
 	}
 
-	// For local estimator, there is endpoint provided, thus we should let
-	// model component decide whether/how to init
-	model.CreatePowerEstimatorModels(
-		collector_metric.ContainerFeaturesNames,
-		collector_metric.NodeMetadataFeatureNames,
-		collector_metric.NodeMetadataFeatureValues,
-	)
-
 	m := manager.New()
 	prometheus.MustRegister(version.NewCollector("kepler_exporter"))
 	prometheus.MustRegister(m.PrometheusCollector)

diff --git a/pkg/collector/metric_collector.go b/pkg/collector/metric_collector.go
@@ -23,6 +23,7 @@ import (
 	"github.com/sustainable-computing-io/kepler/pkg/bpfassets/attacher"
 	"github.com/sustainable-computing-io/kepler/pkg/cgroup"
 	"github.com/sustainable-computing-io/kepler/pkg/config"
+	"github.com/sustainable-computing-io/kepler/pkg/model"
 	"github.com/sustainable-computing-io/kepler/pkg/power/accelerator/gpu"
 	"github.com/sustainable-computing-io/kepler/pkg/power/accelerator/qat"
 	"github.com/sustainable-computing-io/kepler/pkg/utils"
@@ -81,6 +82,14 @@ func (c *Collector) Initialize() error {
 		return err
 	}
 
+	// For local estimator, there is endpoint provided, thus we should let
+	// model component decide whether/how to init
+	model.CreatePowerEstimatorModels(
+		collector_metric.ContainerFeaturesNames,
+		collector_metric.NodeMetadataFeatureNames,
+		collector_metric.NodeMetadataFeatureValues,
+	)
+
 	c.prePopulateContainerMetrics(pods)
 	c.updateNodeEnergyMetrics()
 

diff --git a/pkg/model/container_energy.go b/pkg/model/container_energy.go
@@ -86,8 +86,13 @@ func createContainerPowerModelConfig(powerSourceTarget string, containerFeatureN
 				collector_metric.GPU + "_IDLE",    // for idle GPU power consumption
 			}...)
 		} else if powerSourceTarget == config.ContainerPlatformPowerKey {
+			platformUsageMetric := config.CoreUsageMetric
+			if !attacher.HardwareCountersEnabled {
+				// Given that there is no HW counter in  some scenarios (e.g. on VMs), we have to use CPUTime data.
+				platformUsageMetric = config.CPUTime
+			}
 			modelConfig.ContainerFeatureNames = []string{
-				config.CoreUsageMetric, // for PLATFORM resource usage
+				platformUsageMetric, // for PLATFORM resource usage
 			}
 			modelConfig.NodeFeatureNames = modelConfig.ContainerFeatureNames
 			modelConfig.NodeFeatureNames = append(modelConfig.NodeFeatureNames, []string{
@@ -107,6 +112,7 @@ func CreateContainerPowerEstimatorModel(containerFeatureNames, systemMetaDataFea
 	ContainerPlatformPowerModel, err = createPowerModelEstimator(modelConfig)
 	if err == nil {
 		klog.Infof("Using the %s Power Model to estimate Container Platform Power", modelConfig.ModelType.String()+"/"+modelConfig.ModelOutputType.String())
+		klog.Infof("Container feature names: %v", modelConfig.ContainerFeatureNames)
 	} else {
 		klog.Infof("Failed to create %s Power Model to estimate Container Platform Power: %v\n", modelConfig.ModelType.String()+"/"+modelConfig.ModelOutputType.String(), err)
 	}
@@ -116,6 +122,7 @@ func CreateContainerPowerEstimatorModel(containerFeatureNames, systemMetaDataFea
 	ContainerComponentPowerModel, err = createPowerModelEstimator(modelConfig)
 	if err == nil {
 		klog.Infof("Using the %s Power Model to estimate Container Component Power", modelConfig.ModelType.String()+"/"+modelConfig.ModelOutputType.String())
+		klog.Infof("Container feature names: %v", modelConfig.ContainerFeatureNames)
 	} else {
 		klog.Infof("Failed to create %s Power Model to estimate Container Component Power: %v\n", modelConfig.ModelType.String()+"/"+modelConfig.ModelOutputType.String(), err)
 	}

diff --git a/pkg/model/process_power.go b/pkg/model/process_power.go
@@ -85,8 +85,13 @@ func createProcessPowerModelConfig(powerSourceTarget string, processFeatureNames
 				collector_metric.GPU + "_IDLE",    // for idle GPU power consumption
 			}...)
 		} else if powerSourceTarget == config.ProcessPlatformPowerKey {
+			platformUsageMetric := config.CoreUsageMetric
+			if !attacher.HardwareCountersEnabled {
+				// Given that there is no HW counter in  some scenarios (e.g. on VMs), we have to use CPUTime data.
+				platformUsageMetric = config.CPUTime
+			}
 			modelConfig.ContainerFeatureNames = []string{
-				config.CoreUsageMetric, // for PLATFORM resource usage
+				platformUsageMetric, // for PLATFORM resource usage
 			}
 			modelConfig.NodeFeatureNames = modelConfig.ContainerFeatureNames
 			modelConfig.NodeFeatureNames = append(modelConfig.NodeFeatureNames, []string{
@@ -106,6 +111,7 @@ func CreateProcessPowerEstimatorModel(processFeatureNames, systemMetaDataFeature
 	ProcessPlatformPowerModel, err = createPowerModelEstimator(modelConfig)
 	if err == nil {
 		klog.Infof("Using the %s Power Model to estimate Process Platform Power", modelConfig.ModelType.String()+"/"+modelConfig.ModelOutputType.String())
+		klog.Infof("Container feature names: %v", modelConfig.ContainerFeatureNames)
 	} else {
 		klog.Infof("Failed to create %s Power Model to estimate Process Platform Power: %v\n", modelConfig.ModelType.String()+"/"+modelConfig.ModelOutputType.String(), err)
 	}
@@ -115,6 +121,7 @@ func CreateProcessPowerEstimatorModel(processFeatureNames, systemMetaDataFeature
 	ProcessComponentPowerModel, err = createPowerModelEstimator(modelConfig)
 	if err == nil {
 		klog.Infof("Using the %s Power Model to estimate Process Component Power", modelConfig.ModelType.String()+"/"+modelConfig.ModelOutputType.String())
+		klog.Infof("Container feature names: %v", modelConfig.ContainerFeatureNames)
 	} else {
 		klog.Infof("Failed to create %s Power Model to estimate Process Component Power: %v\n", modelConfig.ModelType.String()+"/"+modelConfig.ModelOutputType.String(), err)
 	}