The student will explore non-linear machine learning techniques (e.g., decision trees, random forests, gradient boosting, neural networks) and time-series forecasting methods to predict CPU/GPU utilization, memory, I/O, and energy consumption. The work will focus on building models that generalize across pipeline steps and scales, while remaining interpretable and efficient to deploy in SIM-PIPE. Work to be done:
- Background Research
- Review existing approaches for resource and energy prediction in HPC, cloud, and AI pipelines.
- Analyze current SIM-PIPE linear regression implementation and identify its limitations.
- Data Collection
- Run representative pipelines (ETL, ML training, streaming analytics) at different scales on heterogeneous hardware.
- Collect step-level metrics: CPU/GPU utilization, memory, I/O, network, and energy consumption.
- Feature Engineering
- Derive predictive features from pipeline metadata (step type, data size, parallelism, execution duration).
- Explore combining step-level traces with hardware characteristics.
- Model Development
- Implement non-linear predictive models (e.g., random forest, XGBoost, LSTM for time- series).
- Compare against the current linear regression baseline in SIM-PIPE.
- Evaluation
- Metrics: prediction accuracy (RMSE, MAPE), scalability, training cost, inference overhead.
- Compare performance across pipeline types and hardware configurations.
- Integration into SIM-PIPE
- Implement a predictive module that automatically selects the best model for given workload types.
- Provide “what-if” analysis features (e.g., predict energy usage if pipeline runs on GPU vs. CPU).
- Analysis & Guidelines
- Assess trade-offs between model complexity, interpretability, and runtime cost.
- Provide recommendations on model selection for practitioners.