TornadoVMPulse

TornadoVMPulse is an interactive profilling tool for visualizing and analyzing profiling data from TornadoVM. It provides deep insights into kernel execution, data transfer, memory, and power usage, with a modern, user-friendly interface.

Features

• Upload and Analyze Logs: Supports TornadoVM profiler logs in JSON, log, or CSV format.
• Automatic Conversion: Raw logs are automatically converted to CSV for analysis.
• Interactive Visualizations:
  • Task Time Distribution (color-coded by task graph)
  • Memory Usage
  • Power Usage
  • Copy-In/Copy-Out Analysis
  • Total Execution Time Analysis (table, stacked bar, and sunburst chart)
  • Task Graph Analysis (per-task breakdown for a selected task graph)
• Sidebar Configuration: Select which metrics to display and the time unit (ns, ms, sec).
• Downloadable Results: Download generated CSVs after processing.
• Modern UI: Custom logo, collapsible info panels, and beautiful charts.

Usage

1. Install dependencies: bash pip install -r requirements.txt

2. Run the dashboard: bash streamlit run src/app.py

3. Upload a profiling log file:

   • Drag and drop a raw JSON/log or CSV file into the uploader.
   • If a raw file is uploaded, it will be converted to CSV and available for download.

4. Explore the dashboard:

   • Use the sidebar to select which metrics to display and the time unit.
   • Review the visualizations and download the generated CSV if needed.

Visualizations & Metrics

Task Time Distribution

• Bar chart showing the sum of kernel execution times for each task, color-coded by task graph.
• Pie chart for overall distribution.

Example: Task Time Distribution as a Pie Chart (proportion of total kernel time per task):

Example: Task Time Distribution as a Bar Chart (kernel time per task, color-coded by task graph):

Memory Usage

• Bar chart of total and average memory usage (copy-in bytes) for each task graph.

Power Usage

• Bar chart of average power usage for each task graph.

Copy-In/Copy-Out Analysis

• Table and bar chart showing time and number of data transfer operations, and their contribution to total execution time.

Total Execution Time Analysis

• Table breaking down total execution time by task graph and action (kernel, dispatch, copy-in, copy-out), with percentages.
• Stacked bar chart for visual comparison.
• Sunburst chart: Hierarchical breakdown of total execution time by task graph and component, with interactive tooltips.

Example Table, Stacked Bar, and Sunburst Chart:

Task Graph Analysis

• Dropdown to select a specific task graph.
• Stacked bar chart showing how time is spent (kernel, copy-in, copy-out, dispatch) for each task in the selected graph.
• Memory usage per task (if available).

Summary Statistics

• Statistical summary of the selected metrics.

Raw Data

• View the raw input data as a DataFrame.

Sidebar Controls

• Metrics Configuration: Select which visualizations to display.
• Select time unit: Choose between nanoseconds, milliseconds, or seconds.
• Status: See CSV generation status and download the generated CSV.

About the TornadoVM Profiler

• Official Documentation
• Enable the profiler with --enableProfiler console or silent or via the ExecutionPlan API.
• Profiler output includes kernel, dispatch, copy-in/out, and power metrics.
• Supported power metrics: NVIDIA NVML (OpenCL/PTX) and oneAPI Level Zero SYSMAN (SPIRV).
• Save profiler output with --dumpProfiler <FILENAME>.

Customization

• The dashboard uses a custom logo (static/pulse-logo.png).
• You can further customize the UI and add new visualizations as needed.

License

MIT License