VortexTune: GPGPU Configuration Optimizer

This project provides an automated tool to explore and optimize the performance of the Vortex GPGPU architecture.

By utilizing Optuna, this script runs simulations across various hardware configurations to identify the optimal setup that minimizes execution cycles for a given workload (specifically SPLA benchmarks with MTX inputs).

Features

• Automated Design Space Exploration: Automatically tunes hardware parameters to find the best performance.
• Parallel Execution: Supports running multiple simulation workers concurrently.
• Persistent Storage: Uses SQLite to save study progress, allowing you to pause and resume optimization.
• Detailed Logging: Captures stdout/stderr and execution time for every simulation trial.

Tunable Parameters

The optimizer explores the following hardware configuration space:

Parameter	Description	Values Explored
Cores	Number of cores	1, 2, 4, 8
Clusters	Number of clusters	1, 2, 4
Warps	Warps per core	2, 4, 8, 16, 32, 64
Threads	Threads per warp	1, 2, 4, 8, 16, 32, 64
L2 Cache	L2 Cache status	On (1), Off (0)
L3 Cache	L3 Cache status	On (1), Off (0)

Prerequisites

1. Python 3.x
2. Optuna: Install via pip:

   pip install optuna

3. Vortex & SPLA: You must have the Vortex GPGPU simulator and SPLA benchmark compiled and available on your system.
4. Helper Script: Ensure run_vortex.sh is present in the same directory as the python script and is executable (chmod +x run_vortex.sh).

Usage

Run the config_finder.py script with the required paths to your Vortex installation, SPLA directory, and the input matrix file.

Basic Command

python3 config_finder.py \
  --spla_path /path/to/spla_root \
  --vortex_path /path/to/vortex_root \
  --mtx_path /path/to/input_matrix.mtx

Advanced Options

You can control the number of trials and parallel jobs:

python3 config_finder.py \
  --spla_path /home/user/spla \
  --vortex_path /home/user/vortex \
  --mtx_path /home/user/data/matrix.mtx \
  --batch 100 \
  --jobs 10 \
  --study-name "vortex_optimization_experiment_1"

Arguments

• --spla_path: (Required) Path to the SPLA directory.
• --vortex_path: (Required) Path to the Vortex directory.
• --mtx_path: (Required) Path to the .mtx input file.
• --batch: Number of trials to run (default: 10).
• --jobs: Number of parallel workers (default: 10).
• --study-name: Name of the Optuna study (default: "vortex-optimization2").

Output

1. Console: Displays the best parameters and best execution cycle count found so far.
2. Database: Results are stored in configuration.db (SQLite).
3. Logs: Detailed logs for each simulation are saved in the logs_fixed_cores_clusters/ directory.