Replication Package for 'Data-Driven Multi-Objective Optimization of ML Inference Hardware Configurations for Energy, Performance and Cost'

Data-Driven Multi-Objective Optimization of ML Inference Hardware Configurations for Energy, Performance and Cost Replication Package This repository contains the code, datasets, and metadata required to reproduce the findings presented in the paper: "Data-Driven Multi-Objective Optimization of ML Inference Hardware Configurations for Energy, Performance and Cost", submitted to Sustainable Computing: Informatics and Systems (SUSCOM). This framework leverages MLPerf™ Inference (v4.1 and v5.1) data to build a recommender system that identifies Pareto-optimal hardware configurations, balancing throughput, energy efficiency, and hardware cost. 📂 Repository Structure code/ Dataset_Extension.ipynb: Step 1: Data harmonization & enrichment pipeline Data_Analysis.ipynb: Step 2: EDA, correlation analysis, and sensitivity checks (RQ1) Optimization_Model.ipynb: Step 3: Predictive modeling & Pareto optimization (RQ2) paper_metrics.ipynb: Step 4: Reproduces exact abstract/results statistics datasets/ Inference_data_v4.1.csv: Raw input data (v4.1) Inference_data_v5.1.csv: Raw input data (v5.1) Inference_data_Harmonized_Merged.csv: Final enriched dataset (Output of Step 1) metadata/ accelerators.yaml: Specifications (TDP, Memory, Pricing) for GPUs/ASICs cpus.yaml: Specifications for Host CPUs models.yaml: Task domain mappings for MLPerf models figures/: Generated plots (PDFs) from analysis scripts logs/: Generated textual logs and trained models requirements.txt: Python dependencies