How yyysjz1997/sgdp works
Project type and purpose
• This is a research code demo for “SGDP: A Stream-Graph Neural Network Based Data Prefetcher.”
• It is not a user-facing full-stack app, but a standalone Python framework that ingests block I/O traces, builds stream-graph representations, trains/tests a graph neural network model, and simulates a cache to measure prefetching performance.
File structure
• Top-level files:
• cache.py – cache simulator class
• cachetest.py – entry point for cache tests (single-step & rolling prefetching)
• data_process.py – trace ingestion and stream conversion for training/testing
• main.py – orchestrates end-to-end training and evaluation based on arguments
• model.py – defines the SGDP graph neural network, training, testing, prediction routines
• pred_list_gene.py – builds local/global/hybrid connection matrices, produces prefetch lists, computes HR/EPR
• utils.py – shared routines (I/O, graph building helpers, metric functions)
• readme.md – project overview, instructions
• cachetest.py & cache.py – simulate how predicted addresses are fed to the cache
README for additional insights
• Datasets are block I/O traces in streaming form
• Workflow phases: data processing → graph construction → model train/test → cache simulation → metric computation
• Key technologies: Python, graph neural networks, gated graph modules, attention mechanism