GreenTune: LLM-Based Hardware-Software Parameter Autotuning of HPC Applications to Improve Performance and Energy Consumption

GreenTune is an autotuner for optimizing hardware and software parameters of large-scale high-performance computing (HPC) applications. Leveraging OpenAI's GPT-4 language model, GreenTune iteratively refines parameter configurations to balance execution time and energy consumption, enabling sustainable and efficient computing.

Key Features

• Energy-Aware Optimization: GreenTune integrates energy consumption as a key metric alongside execution time, making it ideal for environmentally conscious HPC applications.
• LLM-Driven Tuning: Utilizes state-of-the-art language models to generate optimized parameter configurations based on performance feedback.
• Customizable Parameters: Supports tuning for a wide range of hardware and software configurations.
• Iterative Refinement: Continuously improves performance metrics over multiple iterations.

System Overview

GreenTune consists of the following modules:

• Prompt Generator: Generates optimization prompts for the LLM, incorporating both energy and execution time metrics.
• Option Evaluator: Parses LLM responses to extract new parameter configurations.
• Performance Analyzer: Measures execution time and energy consumption for given configurations.
• Safety Enforcer: Ensures generated parameters are within safe and valid bounds.

Usage

1. Set up your environment by configuring the necessary parameters in config.py.

2. Run the main script using the command:

   python main.py

3. Monitor the iterative optimization process, which outputs the best configurations based on execution time and energy consumption.

Prerequisites

• Python 3.8 or higher
• Access to OpenAI GPT-4 API
• Intel RAPL support for energy measurement (optional but recommended)

Installation

Clone the repository and install dependencies:

git clone https://github.com/yourusername/greentune.git
cd greentune
python3 main.py

Future Work

GreenTune aims to integrate support for additional energy measurement tools and expand to other domains beyond HPC applications.