Skip to content

kepler-release-0.7

Past due by 6 months 60% complete

Road map for the release 0.7

  • Power model training pipeline
  • Train power model from various machines
    • SPECPower: [plus] large amount (>900), [minus] offline dataset with simple metric (CPU utilization)
    • AWS self-hosted instance: [plus] limited number of profiles (40), [minus] on-release kepler metrics (CPU, cache, and so on)
  • Model selection logic with mac…

Road map for the release 0.7

  • Power model training pipeline
  • Train power model from various machines
    • SPECPower: [plus] large amount (>900), [minus] offline dataset with simple metric (CPU utilization)
    • AWS self-hosted instance: [plus] limited number of profiles (40), [minus] on-release kepler metrics (CPU, cache, and so on)
  • Model selection logic with machine similarity report
  • Power model continuous integration and delivery framework
  • Regression testing

Local XGBoost estimator

  • Migrate xgboost to the main pipeline abstract (removal of XGBoostRegressionStandalonePipeline)
  • Train the model and export weight
  • Test xgboost kepler integration with provided weight

Page cache

  • Collect data with page cache hit
  • Update feature group and retrain the model

Model accuracy

  • Add MAPE and change error key to MAPE
  • Improve training with feature relevance knowledge (such as not using memory feature for core model training)

Integration and Deployment

  • Training CI - Tekton
  • Platform validation CI
  • Operator integration
Loading