Workload Analyzer 📊

A simple tool for monitoring and analyzing GPU and system resource usage during AI/ML workloads.

🌟 Features

• 📈 Real-time monitoring of GPU utilization, memory usage, and system resources
• 🔍 Detailed visualization of resource usage over time
• 💡 Intelligent recommendations for workload optimization
• 🚀 Easy-to-use CLI interface
• 📝 Comprehensive statistics export

🔧 Installation

Using pip

pip install workload-analyzer

Development setup

# Clone the repository
git clone https://github.com/yourusername/workload-analyzer
cd workload-analyzer

# Install using uv with development dependencies
uv sync --dev

📋 Requirements

• Python 3.12+
• NVIDIA GPU with nvidia-smi (for GPU monitoring)

🚀 Quick Start

# Run a command with default settings
workload-analyzer "python train_model.py"

# Specify timeout and polling interval
workload-analyzer "python train_model.py" --timeout 300 --interval 5

📊 Output

The tool generates:

1. Statistics file: JSON format data with all recorded measurements

2. Visualizations:

   • GPU memory usage over time
   • System memory consumption
   • CPU and disk utilization
   • Network usage
   • Process memory statistics

3. Optimization recommendations based on resource utilization patterns:

   • GPU memory sizing recommendations
   • Compute utilization insights
   • Memory bandwidth analysis
   • System resource optimization tips

All outputs are saved to workload_results/ by default (configurable with --output-dir).

🛠️ Configuration options

--timeout           Time to monitor in seconds (default: 120)
--interval          Polling interval in seconds (default: 3)
--recommendations   Enable workload optimization recommendations (default: True)
--output-dir        Directory to save outputs (default: workload_results/)
--verbose           Enable verbose logging (default: True)
--version           Print version information

❕ License

Package is licensed under Apache 2.0 license. Free to use as you like, but a cite of the package is welcome:

@misc{skafte_workload_analyzer,
    author       = {Nicki Skafte Detlefsen},
    title        = {Workload-Analyzer},
    howpublished = {\url{https://github.com/SkafteNicki/workload_analyzer}},
    year         = {2025}
}