MARVELL-PSC: Phase-Conditioned ML Storage Controller

Predictive storage tiering and prefetching using Phase-Conditioned Machine Learning.

Project Structure

• psc/: Core package.
  • core/: Simulator engine, tier configuration, and trace loading.
  • models/: ML models (Tiered Predictor, SLM Predictor, RL Controller).
  • integration/: Integration scripts for analytical simulators (e.g., AstraSim).
• scripts/: CLI tools for running simulations and benchmarks.
• data/: Sample trace files and datasets.
• weights/: Trained model checkpoints.
• research/: Archived development scripts and experimental notes.

Getting Started

Installation

pip install -r simulator/requirements.txt

Running a Simulation

Run the main simulation harness with default settings (Zipfian trace):

python scripts/main_sim.py

Run with trained models:

python scripts/main_sim.py --lstm weights/phase_classifier.pth --mlp weights/policy_model_conditioned.pth

Features

• Phase-Conditioned Prediction: Uses an LSTM to classify I/O phases and an MLP to make tiering decisions.
• Multi-Tier Support: Simulates HBM, CXL DRAM, NVMe SSD, and Cold Storage.
• AstraSim Integration: Bridges ML-based decisions with cycle-accurate architectural simulation.