Kernel Benchmark Visualizer

Kernel Benchmark Visualizer is a web-based dashboard for visualizing and analyzing the performance of General Matrix Multiplication (GEMM) kernels. It provides interactive visualizations including a Roofline plot, performance heatmap, and size comparisons of different GEMM kernels.

Features

• Interactive Roofline Plot
• Performance Heatmap
• Kernel Size Comparison Charts
• Filtering by batch size, data type, and model
• Kubernetes-ready deployment

Prerequisites

• Python 3.9+
• Docker
• Kubernetes cluster (for deployment)
• kubectl

Directory Structure

gemm-kernel-visualizer/
│
├── app/
│   ├── gemm_kernel_visualizer.py
│   └── requirements.txt
│
├── kubernetes/
│   └── deployment.yaml
│
├── Dockerfile
│
└── README.md

Installation

1. Clone the repository:

   git clone https://github.com/yourusername/gemm-kernel-visualizer.git
   cd gemm-kernel-visualizer
   

2. Install the required Python packages:

   pip install -r app/requirements.txt
   

Local Usage

To run the application locally:

1. Navigate to the app directory:

   cd app
   

2. Run the Python script:

   python gemm_kernel_visualizer.py
   

3. Open a web browser and go to http://127.0.0.1:8050/ to view the dashboard.

Docker Build

To build the Docker image:

1. Ensure you're in the root directory of the project.

2. Build the image:

   docker build -t gemm-kernel-visualizer:latest .
   

3. Run the container locally (optional):

   docker run -p 8050:8050 gemm-kernel-visualizer:latest
   

Kubernetes Deployment

To deploy the application to a Kubernetes cluster:

1. Push the Docker image to a container registry:

   docker tag gemm-kernel-visualizer:latest your-registry/gemm-kernel-visualizer:latest
   docker push your-registry/gemm-kernel-visualizer:latest
   

2. Update the kubernetes/deployment.yaml file with your image name.

3. Apply the Kubernetes deployment:

   kubectl apply -f kubernetes/deployment.yaml
   

4. Check the deployment status:

   kubectl get deployments
   kubectl get pods
   kubectl get services
   

5. Access the application:

   • If using Minikube: minikube service gemm-kernel-visualizer-service
   • If using a cloud provider, check the external IP of the LoadBalancer service

Customization

To customize the visualizations or add new features:

1. Modify the app/gemm_kernel_visualizer.py file.
2. Update the app/requirements.txt if you add new dependencies.
3. Rebuild the Docker image and redeploy if necessary.

Interactive Plots

Roofline Plot

The Roofline Plot shows:

• Scatter points representing individual kernels
• X-axis: Arithmetic Intensity (FLOP/byte)
• Y-axis: Performance (TFLOP/s)
• Red line: Roofline model combining memory bandwidth and peak compute limits

Interactivity:

• Hover over points to see kernel details
• Use the toolbar to zoom, pan, or reset the view

Performance Heatmap

The Heatmap displays:

• Kernels on the X-axis
• Performance intensity represented by color

Interactivity:

• Hover over cells to see exact performance values
• Use the color scale to identify high and low performing kernels

Split Charts

The Split Charts show:

1. Arithmetic Intensity Chart:

   • X-axis: Kernels
   • Y-axis: Arithmetic Intensity (FLOP/byte)

2. Performance Chart:

   • X-axis: Kernels
   • Y-axis: Performance (TFLOP/s)

Interactivity:

• Hover over bars to see exact values
• Compare arithmetic intensity and performance side-by-side

Filtering Data

Use the dropdown menus at the top of the page to filter the data:

• Batch Size
• Data Type (dtype)
• Model

All charts will update automatically based on your selection.

Customization

To use your own data:

1. Replace the data list in the script with your benchmark results.
2. Adjust the peak_memory_bandwidth and peak_compute variables in the script to match your hardware specifications.

Troubleshooting

• If the plots don't render, check your internet connection (Plotly requires internet access for some components).
• For Windows users, if you encounter issues with auto-reloading, modify the last line of the script to:

app.run_server(debug=True, use_reloader=False)

Contributing

Contributions to improve the visualizer are welcome! Please submit issues and pull requests on our GitHub repository.

License

This project is licensed under the MIT License - see the LICENSE file for details.